A Review of 40 Years in Cognitive Architecture Research

Core Cognitive Abilities and Practical Applications

Iuliia Kotseruba · John K. Tsotsos

arXiv:1610.08602v3 [cs.AI] 13 Jan 2018

Abstract In this paper we present a broad overview of the last 40 years of research on cognitive architectures.
Although the number of existing architectures is nearing several hundred, most of the existing surveys do not
reflect this growth and focus on a handful of well-established architectures. Thus, in this survey we wanted
to shift the focus towards a more inclusive and high-level overview of the research on cognitive architectures.
Our final set of 84 architectures includes 49 that are still actively developed, and borrow from a diverse set
of disciplines, spanning areas from psychoanalysis to neuroscience. To keep the length of this paper within
reasonable limits we discuss only the core cognitive abilities, such as perception, attention mechanisms,
action selection, memory, learning and reasoning. In order to assess the breadth of practical applications
of cognitive architectures we gathered information on over 900 practical projects implemented using the
cognitive architectures in our list.
We use various visualization techniques to highlight overall trends in the development of the field. In
addition to summarizing the current state-of-the-art in the cognitive architecture research, this survey de-
scribes a variety of methods and ideas that have been tried and their relative success in modeling human
cognitive abilities, as well as which aspects of cognitive behavior need more research with respect to their
mechanistic counterparts and thus can further inform how cognitive science might progress.
Keywords Survey · Cognitive architectures · Perception · Attention · Cognitive abilities · Practical
applications

1 Introduction

The goal of this paper is to provide a broad overview of the last 40 years of research in cognitive architectures
with an emphasis on the core capabilities of perception, attention mechanisms, action selection, learning,
memory, reasoning and their practical applications. Although the field of cognitive architectures has been
steadily expanding, most of the surveys published in the past 10 years do not reflect this growth and
feature essentially the same set of a dozen most established architectures. The latest large-scale study was

Iuliia Kotseruba
E-mail: yulia k@cse.yorku.ca
John K. Tsotsos
E-mail: tsotsos@cse.yorku.ca
2 Iuliia Kotseruba, John K. Tsotsos

conducted by Samsonovich in 2010 [456] in an attempt to catalog the implemented cognitive architectures.
His survey contains descriptions of 26 cognitive architectures submitted by their respective authors. The
same information is also presented online in a Comparative Table of Cognitive Architectures1 . In addition to
surveys, there are multiple on-line sources listing cognitive architectures, but they rarely go beyond a short
description and a link to the project site or a software repository.
Since there is no exhaustive list of cognitive architectures, their exact number is unknown, but it is
estimated to be around three hundred, out of which at least one-third of the projects are currently active. To
form the initial list for our study we combined the architectures mentioned in surveys (published within the
last 10 years) and several large on-line catalogs. We also included more recent projects not yet mentioned
in the survey literature. Figure 1 shows a visualization of 195 cognitive architectures featured in 17 sources
(surveys, on-line catalogs and Google Scholar). It is apparent from this diagram that a small group of
architectures such as ACT-R, Soar, CLARION, ICARUS, EPIC, LIDA and a few others are present in most
sources, while all other projects are only briefly mentioned in on-line catalogs. While the theoretical and
practical contributions of the major architectures are undeniable, they represent only a part of the research
in the field. Thus, in this review the focus is shifted away from the deep study of the major architectures
or discussing what could be the best approach to modeling cognition, which has been done elsewhere. For
example, in a recent paper by Laird et al. [303] ACT-R, Soar and Sigma are compared based on their
structural organization and approaches to modelling core cognitive abilities. Further, a new Standard Model
of the Mind is proposed as a reference model born out of consensus between the three architectures. Our
goal, on the other hand, is to present a broad, inclusive, judgment-neutral snapshot of the past 40 years of
development with the goal of informing the future cognitive architecture research by presenting the diversity
of ideas that have been tried and their relative success.
To make this survey manageable we reduced the original list of architectures to 84 items by considering
only implemented architectures with at least one practical application and several peer-reviewed publica-
tions. Even though we do not explicitly include some of the philosophical architectures such as CogAff [492],
Society of Mind [369], Global Workspace Theory (GWT) [28] and Pandemonium theory [480], we exam-
ine cognitive architectures heavily influenced by these theories (e.g. LIDA, ARCADIA, CERA-CRANIUM,
ASMO, COGNET and Copycat/Metacat, also see discussion in Section 5). We also exclude large-scale brain
modeling projects, which are low-level and do not easily map onto the breadth of cognitive capabilities
modeled by other types of cognitive architectures. Further, many of the brain models do not have practical
applications, and thus do not fit the parameters of the present survey. Figure 2 shows all architectures fea-
tured in this survey with their approximate timelines recovered from the publications. Of these projects 49
are currently active2 .
As we mentioned earlier, the first step towards creating an inclusive and organized catalog of imple-
mented cognitive architectures was made by Samsonovich [456]. His work contained extended descriptions of
26 projects with the following information: short overview, a schematic diagram of major elements, common
components and features (memory types, attention, consciousness, etc.), learning and cognitive development,
cognitive modeling and applications, scalability and limitations. A survey of this kind brings together re-
searchers from several disjoint communities and helps to establish a mapping between the different approaches
and terminology they use. However, the descriptive or tabular format does not allow easy comparisons be-

1
http://bicasociety.org/cogarch/architectures.htm
2
Since the exact dates for the start (and end) of the development are not specified for the majority of architectures, we use
instead the first and the latest publication. If an architecture has an associated website or an online repository, we consider
the project currently active, given that there was any activity (site update/code commit) within the last year. Similarly, we
consider the project under the active development if there was at least one publication within the last year (2016).
A Review of 40 Years in Cognitive Architecture Research Core Cognitive Abilities and Practical Applications 3

2 0 15

2 0 15
2

2
G se l 1 2 2 0 1

G se l 1 2 2 0 1
w ert h 0 1 4

As k 2 son 010

12

w ert h 0 1 4

As k 2 son 010

12
d e s 20 0 1

d e s 20 0 1
Ve fan 00 06

Ve fan 00 06
el 20

el 20
01

01
Th Ga 07

Th Ga 07
o li ty

a n 2

o li ty

a n 2
o 2 0

o 2 0
8

8
G gar el 2

L a i el 2

G o r is is 2

G gar el 2

L a i el 2

G o r is is 2
c o ert n k s

c o ert n k s
rtz an

rtz an
Pr ch y 2

Pr ch y 2
G ch - ci e

G ch - ci e
lin o r

lin o r
e

e
oe m

oe m
t

0
r

r
z
o c
ik z

u e

z
o c
ik z

u e
d u a so

d u a so
D n gl

D ngl
n

n
r

r
o

o
c

c
GOOGLE SCHOLAR:

bi

bi
ACT-R 3T
CLARION AKIRA ARCADIA
Soar APEX CECEP
EPIC ASMO CHRIS
LIDA Architecture of Brain and Mind Casimir
ICARUS BioRC DAC
4CAPS Brain-Based Devices ECA
DUAL Brain-i-Nets EmoCog
OpenCog BrainScales HiTECH
Polyscheme C2S2 IOCA
NARS CIRCA Ikaros
4D/RCS CMattie MACSi
ART CoJACK MUSICOG
GLAIR CogAff MuRCA
SNePS CogNet PERSEED
Shruti Cognitive Computation Project RoboCog
ADAPT Companions SBR
AIS Cortronics SOIMA
CHREST DISCERN and PROC The Meta-Cognitive Loop
CogPrime DiPRA robo-CAMAL
Copycat/Metacat Digital Aristotle CARACaS
DeSTIN Disciple CHARISMA
ERE EPAM CTRNN
Global Workspace Theory Emergent CoSy
HTM Emile DSO
IBCA Evolved Machines ERA
MicroPsi FACETS Haikonen CA
NOMAD Gat IMA
PRODIGY Guardian ITALK
PRS H-Cogaff
STAR
SAIL Hal
SAL IBM Watson ARDIS
AAR IMPRINT CELTS
ATLANTIS Ikon Flux CRAM
BECCA Large-Scale Brain Modeling Cognitive Symmetry Engine
Blue Brain MAMID Darwinian Neurodynamics
CERA-CRANIUM Mibe DIARC
CYC NCS EFAA
Cerebus NEAT and HyperNEAT Feelix Growing
DAV NWS IAC
Darwin Nengo ISAC
Emotion Machine Neocortex JDE
FLOWERS Nexting MDB
FORR Novamente Meta-AQUA
GMU-BICA NuPIC NEUCOGAR
HUMANOID Numenta Psikharpax
Homer OMAR Roillo
I-C SDAL OpenNERO SERA
IM-CLEVER Parallel terraced scan SiMA
Kismet Pogamut SPA
The Ouroboros Model
LEABRA PolyScheme Joshua Blue
MAX Psi-Theory MLECOG
Neurogrid RCS MoNAD
OSCAR REM NIMBUS
PAL Recommendation Architecture ROSSI
R-CAST SILK SACA
RALPH-MEA Sigma SMRITI
SASE Society of Mind Storm
Subsumption SpiNNaker VICAL
Teton TexAI Xapagy
Theo Thalamo-Cortical Systems
Ymir The Ersatz Brain Project
The Remote Agent
The Society of Mind
TinyCog
Tosca
YKY project
iCub
REFERENCES:
bicasociety http://bicasociety.org/cogarch/architectures.htm
duch-links https://duch-links.wikispaces.com/Cogni tive+Architectures Vernon 2007 Vernon, D. et al. (2007 ). A Survey of Artificial Cogni tive Systems Implications
Goertzel 2014 Goertzel, B. (2014 ). Brief Survery Of Cogni tive Architectures. for the Autonomous Develop ment of Mental Capabili ties in ComputationaAgents. l
Enginee ring General Intelligen ce, Part 1.Atlantis Press, 2014 . 101-142. IEEE Transactions on Evolutionary Computation, 1–30 .
cogarch http://cogarch.org/inde x.php/Main_ Page linas http://lina s.org/agi.html

Goertzel 2010 Goertzel, B. et al. (2010 ). A world survey of artificial brain projects, de Garis 2010 de Garis et al., (2010 ). A world survey of artificial brain projects,
Part II: Biologi cally inspired cogni tive architectures. Neurocomputing, 74(1-3), 30–49 . Part I: Large-scale brain simulations. Neurocomputing, 74(1-3), 3–29 .
wiki https://en.wikipedia .org/wiki/Comparison_o f_cogni tive_architectures Thorisson 2012 Thórisson, K.,Helga sson, H. (2012 ). Cogni tiveArchitectures and Autonomy:
A Comparative Review. Journal of Artificial General Intelligen ce, 3(2), 1–30
Langley 2006 Langl ey, P. et al. (2006 ). Cogni tive architectures: Research issues and challenge s.
Cogni tive Systems Research, 10, 141–160 . Gok 2012 Gök, S. E., Sayan, E. (2012 ). A philo sophi cal assessment of computational model s
of consciousness. Cogni tive Systems Research, 17-18, 49–62 .
Duch 2008 Duch, W. et al. (2008 ). Cogni tive Architectures: Where do we go from here?
AGI-2008 , 171, 122–136 . Asselman 2015 Asselman, A. et al. (2015 ). Comparative Study of Cogni tiveArchitectures. IRJCS, 2(9), 8–13 .
Profanter 2012 Profanter, S. (2012 ). Cogni tive architectures. Haup te Seminar Human Robo t Interaction. Goertzel 2012 Goertzel, B. (2012 ). Perception processing for gene ral intelligen ce: Bridging the
symboli c/subsymboli c gap. Lecture Notes in Computer Science, 79–88 .

Fig. 1 A diagram showing cognitive architectures found in literature surveys and on-line sources shown with blue and orange
colors respectively. The architectures in the diagram are sorted by the total number of references in the surveys and on-line
sources for each architecture. Titles of the architectures covered in this survey are shown in red.
All visualizations in this paper are made using D3.js library (https://d3js.org) and interactive versions of the figures are
available on the project website. The color palettes are generated using ColorBrewer (http://colorbrewer2.org).
4 Iuliia Kotseruba, John K. Tsotsos

tween architectures. Since our sample of architectures is large, we experimented with alternative visualization
strategies.
In the following sections, we will provide an overview of the definitions of cognition and approaches to
categorizing cognitive architectures. As one of our contributions, we map cognitive architectures according
to their perception modality, implemented mechanisms of attention, memory organization, types of learning,
action selection and practical applications.
In the process of preparing this paper, we examined the literature widely and this activity led to a 2500
item bibliography of relevant publications. We provide this bibliography, with short summary descriptions
for each paper as a supplementary material. In addition, interactive versions of all diagrams in this paper
are available on-line3 and allow one to explore the data and view relevant references.

2 What are Cognitive Architectures?

Cognitive architectures are a part of research in general AI, which began in the 1950s with the goal of creating
programs that could reason about problems across different domains, develop insights, adapt to new situations
and reflect on themselves. Similarly, the ultimate goal of research in cognitive architectures is to model the
human mind, which will bring us closer to building human-level artificial intelligence. Cognitive architectures
attempt to provide evidence that particular mechanisms succeed in producing intelligent behavior and thus
contribute to cognitive science. Moreover, the body of work represented by the cognitive architectures, and
this review, documents what methods or strategies have been tried previously (and what have not), how
they have been used, and what level of success has been achieved or lessons learned, all important elements
that help guide future research efforts. For AI and engineering, documentation of past mechanistic work has
obvious import. But this is just as important for cognitive science, since most experimental work eventually
attempts to connect to explanations of how observed human behavior may be generated and the body of
cognitive architectures provides a very rich source of viable ideas and mechanisms.
According to Russel and Norvig [452] artificial intelligence may be realized in four different ways: systems
that think like humans, systems that think rationally, systems that act like humans, and systems that act
rationally. The existing cognitive architectures have explored all four possibilities. For instance, human-like
thought is pursued by the architectures stemming from cognitive modeling. In this case, the errors made
by an intelligent system should match the errors typically made by people in similar situations. This is in
contrast to rationally thinking systems which are required to produce consistent and correct conclusions for
arbitrary tasks. A similar distinction is made for machines that act like humans or act rationally. Machines
in either of these groups are not expected to think like humans, only their actions or behavior is taken into
account.
However, with no clear definition and general theory of cognition, each architecture was based on a dif-
ferent set of premises and assumptions, making comparison and evaluation difficult. Several papers were
published to resolve the uncertainties, the most prominent being Sun’s desiderata for cognitive architectures
[505] and Newell’s functional criteria (first published in [386] and [387], and later restated by Anderson and
Lebiere in [17]). Newell’s criteria include flexible behavior, real-time operation, rationality, large knowledge
base, learning, development, linguistic abilities, self-awareness and brain realization. Sun’s desiderata are
broader and include ecological, cognitive and bio-evolutionary realism, adaptation, modularity, routineness
and synergistic interaction. Besides defining these criteria and applying them to a range of cognitive architec-
tures, Sun also pointed out the lack of clearly defined cognitive assumptions and methodological approaches,
3 http://jtl.lassonde.yorku.ca/project/cognitive_architectures_survey/
A Review of 40 Years in Cognitive Architecture Research Core Cognitive Abilities and Practical Applications 5

Darwinian Neurodynamics
ARCADIA
CORTEX
MLECOG
NEUCOGAR
STAR
RoboCog
SPA
MusiCog
CSE
MIDCA
DSO
CELTS
CHARISMA
MACSi
ASMO
CERA-CRANIUM
CogPrime
Sigma
Xapagy
SAL
BECCA
ARDIS
DiPRA
CARACaS
LIDA
Pogamut
DIARC
HTM
GMU-BICA
ADAPT
ARS/SiMA
CoJACK
Companions
iCub
CoSy
MicroPsi
PolyScheme
MDB
Novamente
R-CAST
MAMID
REM
SASE
Kismet
APEX
CLARION
IMPRINT
Leabra
3T
Ymir
EPIC
DUAL
OMAR
Shruti
CHREST
DAC
GLAIR
FORR
ATLANTIS
ISAC
Recommendation
ERE
COGNET
ICARUS
MAX
Teton
Theo
RALPH
Disciple
Homer
OSCAR
PRODIGY
PRS
MIDAS
Subsumption
Copycat/Metacat
NARS
AIS
Soar
CAPS
BBD
ACT-R
ART
RCS

1975 1980 1985 1990 1995 2000 2005 2010 2015

Fig. 2 A timeline of 84 cognitive architectures featured in this survey. Each line corresponds to a single architecture. The
architectures are sorted by the starting date, so that the earliest architectures are plotted at the bottom of the figure. Since
the explicit beginning and ending dates are known only for a few projects, we recovered the timeline based on the dates of the
publications and activity on the project web page or on-line repository. Colors correspond to different types of architectures:
symbolic (green), emergent (red) and hybrid (blue). According to this data there was a particular interest in symbolic architec-
tures since mid-1980s until early 1990s, however after 2000s most of the newly developed architectures are hybrid. Emergent
architectures, many of which are biologically-inspired, are distributed fairly evenly in the timeline, but remain a relatively small
group.
6 Iuliia Kotseruba, John K. Tsotsos

which hinder progress in studying intelligence. He also noted an uncertainty regarding essential dichotomies
(implicit/explicit, procedural/declarative, etc.), modularity of cognition and structure of memory. However,
a quick look at the existing cognitive architectures reveals persisting disagreements in terms of their research
goals, structure, operation and application.
Instead of looking for a particular definition of intelligence [323], it may be more practical to define it as
a set of competencies and behaviors demonstrated by the system. While no comprehensive list of capabilities
required for intelligence exists, several broad areas have been identified that may serve as guidance for ongoing
work in the cognitive architecture domain. For example, Adams et al.[1] suggest areas such as perception,
memory, attention, actuation, social interaction, planning, motivation, emotion, etc. These are further split
into subareas. Arguably, some of these categories may seem more important than the others and historically
attracted more attention (further discussed in Section 11.2).
Implementing even a reduced set of abilities in a single architecture is a substantial undertaking. Unsur-
prisingly, the goal of achieving Artificial General Intelligence (AGI) is explicitly pursued by a small number
of architectures, among which are Soar, ACT-R, NARS [580], LIDA [150], and several recent projects, such
as SiMA (formerly ARS) [469], Sigma [426] and CogPrime [197]. Others focus on a particular aspect of
cognition, e.g. attention (ARCADIA [72], STAR [549]), emotion (CELTS [151]), perception of symmetry
(Cognitive Symmetry Engine [223]) or problem solving (FORR [136], PRODIGY [136]). There are also nar-
rowly specialized architectures designed for particular applications, such as ARDIS [352] for visual inspection
of surfaces or MusiCog [359] for music comprehension and generation.
Different opinions can be found in the literature on what system can be considered a cognitive architecture.
Laird in [306] discusses how cognitive architectures differ from other intelligent software systems. While all
of them have memory storage, control components, data representation, and input/output devices, the latter
provide only a fixed model for general computation. Cognitive architectures, on the other hand, must change
through development and efficiently use knowledge to perform new tasks. Furthermore, he suggests toolkits
and frameworks for building intelligent agents (e.g. GOMS, BDI, etc.) cannot themselves be considered
cognitive architectures. (However, the authors of Pogamut, a framework for building intelligent agents,
consider it a cognitive architecture as it is included in [456].) Another opinion on the matter is by Sun
[506], who contrasts the engineering approach taken in the field of artificial intelligence with the scientific
approach of cognitive architectures. According to Sun, psychologically based cognitive architectures should
facilitate the study of human mind by modeling not only the human behavior but also the underlying
cognitive processes. Such models, unlike software engineering oriented ”cognitive” architectures, are explicit
representations of the general human cognitive mechanisms, which are essential for understanding of the
mind.
In practice the term ”cognitive architecture” is not as restrictive, as made evident by the representative
surveys of the field. Most of the surveys define cognitive architectures as a blueprint for intelligence, or
more specifically, a proposal about the mental representations and computational procedures that operate
on these representations enabling a range of intelligent behaviors [128,317,530,425,82]. There is generally no
need to justify the inclusion of the established cognitive architectures such as Soar, ACT-R, EPIC, LIDA,
CLARION, ICARUS and a few others. The same applies to many biomimetic and neuroscience-inspired
cognitive architectures that model cognitive processes on a neuronal level (e.g. CAPS, BBD, BECCA, DAC,
SPA). It can be argued that some engineering oriented architectures pursue a similar goal as they have a set
of structural components for perception, reasoning and action, and model interactions between them, which
is in the spirit of the Newell’s call for ”unified theories of cognition” [212]. However, when it comes to less
common or new projects, the reasons for considering them are less clear. As an example, AKIRA, a framework
that explicitly does not self-identify as a cognitive architecture [416], is featured in some surveys anyway
A Review of 40 Years in Cognitive Architecture Research Core Cognitive Abilities and Practical Applications 7

[533]. Similarly, a knowledge base Cyc [167], which does not make any claims about general intelligence, is
presented as an AGI architecture in [194].
Recently, claims have been made that deep learning is capable of solving AI by Google (DeepMind4 ).
Likewise, Facebook AI Research (FAIR5 ) and other companies are actively working in the same direction.
However, the question is where does this work stand with respect to cognitive architectures? Overall, the
DeepMind research addresses a number of important issues in AI, such as natural language understanding,
perceptual processing, general learning, and strategies for evaluating artificial intelligence. Although partic-
ular models already demonstrate cognitive abilities in limited domains, at this point they do not represent
a unified model of intelligence.
Differently from DeepMind, Mikolov et al. from a Facebook research team explicitly discuss their work
in a broader context of developing intelligent machines [366]. Their main argument is that AI is too complex
to be built all at once and instead its general characteristics should be defined first. Two such characteristics
of intelligence are defined, namely, communication and learning, and a concrete roadmap are proposed for
developing them incrementally.
Currently, there are no publications about developing such a system, but overall the research topics
pursued by FAIR align with their proposal for AI and also the business interests of the company. Common
topics include visual processing, especially segmentation and object detection, data mining, natural language
processing, human-computer interaction and network security. Since the current deep learning techniques
are mainly applied to solving practical problems and do not represent a unified framework we do not include
them in this review. However, given their prevalence in other areas of AI, deep learning methods will likely
play some role in the cognitive architectures of the future.
To ensure both inclusiveness and consistency, cognitive architectures in this survey are selected based
on the following criteria: self-evaluation as cognitive, robotic or agent architecture, existing implementation
(not necessarily open-source), and mechanisms for perception, attention, action selection, memory and learn-
ing. Furthermore, we considered the architectures with at least several peer-reviewed papers and practical
applications beyond simple illustrative examples. For the most recent architectures still under development,
some of these conditions were relaxed.
An important point to keep in mind while reading this survey is that cognitive architectures should be
distinguished from the models or agents that implement them. For instance, ACT-R, Soar, HTM and many
other architectures are used as the basis for multiple software agents that demonstrate only a subset of
capabilities declared in theory. On the other hand, some agents may implement extra features that are not
available in the cognitive architecture. A good example is the perceptual system of Rosie [285], one of the
agents implemented in Soar, whereas Soar itself does not include a real perceptual system for physical sensors
as part of the architecture. Unfortunately, in many cases the level of detail presented in the publications does
not allow one to judge whether the particular capability is enabled by the architectural mechanisms (and is
common among all models) or is custom-made for a concrete application. Therefore, to avoid confusion, we
do not make this distinction and list all capabilities demonstrated by the architecture.
Note also that some architectures went through structural and conceptual changes throughout their
development, notably the long-running projects such as Soar, ACT-R, CLARION, Disciple and others. For
example, Soar 9 and subsequent versions added some non-symbolic elements [307] and CAPS4 better accounts
for the time course of cognition and individual differences compared to its predecessors 3CAPS and CAPS6 .

4 https://deepmind.com/
5 https://research.facebook.com/ai/
6 http://www.ccbi.cmu.edu/4CAPS/
8 Iuliia Kotseruba, John K. Tsotsos

Sometimes these changes are well documented (e.g. CLARION7 , Soar8 ), but more often they are not. In
general we use the most recent variant of the architecture for analysis and assume that the features and
abilities of the previous versions are retained in the new version unless there is contradicting evidence.

3 Taxonomies of Cognitive Architectures

Many papers published within the last decade address the problem of evaluation rather than categorization
of cognitive architectures. As mentioned earlier, Newell’s criteria [386,387,17] and Sun’s desiderata [505]
belong in this category. Furthermore, surveys of cognitive architectures propose various capabilities, prop-
erties and evaluation criteria, which include recognition, decision making, perception, prediction, planning,
acting, communication, learning, goal setting, adaptability, generality, autonomy, problem solving, real-time
operation, meta-learning, etc. [568,317,533,26].
While these criteria could be used for classification, many of them are too fine-grained to be applied
to a generic architecture. A more general grouping of architectures is based on the type of representation
and information processing they implement. Three major paradigms are currently recognized: symbolic (also
referred to as cognitivist), emergent (connectionist) and hybrid. Which of these representations, if any,
correctly reflects the human cognitive processes, remains an open question and has been debated for the last
thirty years [509,273].
Symbolic systems represent concepts using symbols that can be manipulated using a predefined instruc-
tion set. Such instructions can be implemented as if-then rules applied to the symbols representing the facts
known about the world (e.g. ACT-R, Soar and other production rule architectures). Because it is a natural
and intuitive representation of knowledge, symbolic manipulation remains very common. Although by de-
sign, symbolic systems excel at planning and reasoning, they are less able to deal with the flexibility and
robustness that are required for dealing with a changing environment and for perceptual processing.
The emergent approach resolves the adaptability and learning issues by building massively parallel models,
analogous to neural networks, where information flow is represented by a propagation of signals from the
input nodes. However, the resulting system also loses its transparency, since knowledge is no longer a set of
symbolic entities and instead is distributed throughout the network. For these reasons, logical inference in a
traditional sense becomes problematic (although not impossible) in emergent architectures.
Naturally, each paradigm has its strengths and weaknesses. For example, any symbolic architecture
requires a lot of work to create an initial knowledge base, but once it is done the architecture is fully
functional. On the other hand, emergent architectures are easier to design, but they must be trained in order
to produce useful behavior. Furthermore, their existing knowledge may deteriorate with the subsequent
learning of new behaviors.
As neither paradigm is capable of addressing all major aspects of cognition, hybrid architectures attempt
to combine elements of both symbolic and emergent approaches. Such systems are the most common in our
selection of architectures (and, likely, overall). In general, there are no restrictions on how the hybridization
is done and many possibilities have been explored. Multiple taxonomies of the hybridization types have
been proposed [504,229,593,594,130]. In addition to representation, one can consider whether the system is
single or multi-module, heterogeneous or homogeneous, take into account the granularity of hybridization
(coarse-grained or fine-grained), the coupling between the symbolic and sub-symbolic components, and types
of memory and learning. In addition, not all the hybrid architectures explicitly address what is referred to as
7 http://www.clarioncognitivearchitecture.com/release-notes
8 https://soar.eecs.umich.edu/articles/articles
A Review of 40 Years in Cognitive Architecture Research Core Cognitive Abilities and Practical Applications 9

EMERGENT HYBRID

connectionist neuronal symbolic

fully integrated
logic systems modeling sub-processing SYMBOLIC
iCub
Ymir
Xapagy
Soar
Sigma
SAL
REM
RALPH
PolyScheme
Novamente
NARS
MLECOG
MIDCA
MACSi
LIDA
Kismet
ISAC
GMU-BICA
FORR
DiPRA
DUAL
DSO
Copycat/Metacat
CogPrime
CoJACK
CSE
CORTEX
CLARION
CHREST
CHARISMA
CERA-CRANIUM
CELTS
CAPS
ASMO
ARS/SiMA
ARCADIA
ADAPT
ACT-R
Subsumption
Shruti
SASE
MDB
BECCA

SPA
Recommendation

STAR
RoboCog
RCS
Pogamut
DIARC
CoSy
CARACaS
MicroPsi
Leabra
HTM
Darwinian Neurodynamics
DAC
BBD
ART

ATLANTIS
ARDIS
3T
Theo
Teton
R-CAST
PRS
PRODIGY
OSCAR
OMAR
MusiCog
MIDAS
MAX
MAMID
IMPRINT
ICARUS
Homer
GLAIR
ERE
EPIC
Disciple
Companions
COGNET
APEX
AIS
Fig. 3 A taxonomy of cognitive architectures based on the representation and processing. The order of the architectures within
each group is alphabetical and does not correspond to the proportion of symbolic vs sub-symbolic elements (i.e. the spatial
proximity of ACT-R and iCub to nodes representing symbolic and emergent architectures respectively does not imply that
ACT-R is closer to symbolic paradigm and iCub is conceptually related to emergent architectures).

symbolic and sub-symbolic elements and reasons for combining them. Only a few architectures, namely ACT-
R, CLARION, DUAL, CogPrime, CAPS, SiMA, GMU-BICA and Sigma, view this integration as essential
and discuss it extensively. However, we found that symbolic and sub-symbolic parts cannot be identified
for all reviewed architectures due to lack of such fine-grained detail in many publications, thus we focus on
representation and processing.
Figure 3 shows the architectures grouped according to the new taxonomy. The top level of the hierarchy is
represented by the symbolic, emergent and hybrid approaches. The definitions of these terms in the literature
are vague, which often leads to the inconsistent assignment of architectures to either group. There seems to
be no agreement even regarding the most known architectures, Soar and ACT-R. Although both combine
symbolic and sub-symbolic elements, ACT-R explicitly self-identifies as hybrid, while Soar does not [307].
The surveys are inconsistent as well, for instance, both Soar and ACT-R are called cognitivist in [568] and
[191], while [26] lists them as hybrid.
The authors of the architectures rarely specify the types of representations used in their systems. Specif-
ically, out of the 84 architectures we surveyed, only 34 have a self-assigned label. Fewer still provide details
on the particular symbolic/subsymbolic processes or elements. It is universally agreed that symbols (labels,
strings of characters, frames), production rules and non-probabilistic logical inference are symbolic and dis-
tributed representations such as neural networks are sub-symbolic. For example, probabilistic action selection
is considered as symbolic in CARACaS [251], CHREST [476] and CogPrime [189], but is described as sub-
symbolic in ACT-R [321], CELTS [151], CoJACK [434], Copycat/Metacat [350] and iCub [459]. Likewise,
numeric data is treated as symbolic in CAPS [266], AIS [212] and EPIC [279], but is regarded as sub-
10 Iuliia Kotseruba, John K. Tsotsos

symbolic in SASE [586]. Similarly, there are disagreements regarding the status of the activations (weights
or probabilities) assigned to the symbols, reinforcement learning, image data, etc.9
To avoid inconsistent grouping, we did not rely on self-assigned labels or conflicting classification of
elements found in the literature. We assume that explicit symbols are atoms of symbolic representation
and can be combined to form meaningful expressions. Such symbols are used for inference or syntactical
parsing. Sub-symbolic representations are generally associated with the metaphor of a neuron. A typical
example of such representation is the neural network, where knowledge is encoded as a numerical pattern
distributed among many simple neuron-like objects. Weights associated with units affect processing and
are acquired by learning. For our classification, we assume that anything that is not an explicit symbol
and processing other than syntactic manipulation is sub-symbolic (e.g. numeric data, pixels, probabilities,
spreading activations, reinforcement learning, etc.). Hybrid representation combines any number of elements
from both representations.
Given these definitions, we assigned labels to all the architectures and visualized them as shown in Figure
3. We distinguish between two groups in the emergent category: neuronal models, which implement models
of biological neurons, and connectionist logic systems, which are closer to artificial neural networks. Within
the hybrid architectures we separate symbolic sub-processing as a type of hybridization where a symbolic
architecture is combined with self-contained module performing sub-symbolic computation, e.g. for processing
sensory data. Other types of functional hybrids also exist, for example, co-processing, meta-processing and
chain-processing, however, these are harder to identify from the limited data available in the publications
and are not included. The fully integrated category combines all other types of hybrids.
The architectures in the symbolic sub-processing group include at least one sub-symbolic module for
sensory processing, while the rest of the knowledge and processing is symbolic. For example, 3T, ATLANTIS,
RCS, DIARC, CARACaS and CoSy are robotic architectures, where a symbolic planning module determines
the behavior of the system and one or more modules are used to process visual and audio sensory data using
techniques like neural networks, optical flow calculation, etc. Similarly, ARDIS and STAR combine symbolic
knowledge base and case-based symbolic reasoning with image processing algorithms.
The fully integrated architectures use a variety of approaches for combining different representations.
ACT-R, Soar, CAPS, Copycat/Metacat, CHREST, CHARISMA, CELTS, CoJACK, CLARION, REM,
NARS and Xapagy combine symbolic concepts and rules with sub-symbolic elements such as activation
values, spreading activation, stochastic selection process, reinforcement learning, etc. On the other hand,
DUAL consists of a large number of highly interconnected hybrid agents, each of which has a symbolic and
sub-symbolic representation, implementing integration at a micro-level. In the case of SiMA hybridization is
done on multiple levels: neural networks are used to build neurosymbols from sensors and actuators (connec-
tionist logic systems) and the top layer is a symbol processing system (symbolic subprocessing). For the rest
of the architectures, it is hard to clearly define hybridization strategy. For instance, many architectures are
implemented as a set of interconnected competing and cooperating modules, where individual modules are
not restricted to a particular representation (Kismet, LIDA, MACSi, ISAC, iCub, GMU-BICA, CORTEX,
ASMO, CELTS, PolyScheme, FORR).
Another hybridization strategy is to combine two different stand-alone architectures with complementary
features. Even though additional effort is required to build interfaces for communication between them, this
approach takes advantage of the strengths of each architecture. A good overview of conceptual and technical
challenges involved in creating an interface between cognitive and robotic architectures is given in [472]. The
9 Activations are considered symbolic only in MusiCog [359], but majority of the architectures list them as sub-symbolic (e.g.

CAPS [266], ACT-R [378, 521], DUAL [290]). Reinforcement learning is considered symbolic in CARACaS [251], CHREST [476]
and MusiCog [359], but sub-symbolic in MAMID [431] and REM [382]. Image data is sub-symbolic in MAMID [431] and SASE
[586] and non-symbolic (iconic) in Soar [307] and RCS [9].
A Review of 40 Years in Cognitive Architecture Research Core Cognitive Abilities and Practical Applications 11

proposed framework is demonstrated with two examples: ACT-R/DIARC and ICARUS/DIARC integration.
Other attempts found in the literature are CERA-CRANIUM/Pogamut [22] and CERA-CRANIUM/Soar
[332] hybrids for playing video games and IMPRINT/ACT-R for human error modeling [322].
Few long-standing projects (based on the number of publications) also exist. A canonical example is
SAL, which comprises ACT-R and Leabra architectures [260]. Here ACT-R is used to guide learning of
Leabra models. In the robotic architecture ADAPT, Soar is utilized for control, whereas separate modules
are responsible for modeling a 3D world from sensor information and for visual processing [48].
Compared to hybrids, emergent architectures form a more uniform group. As mentioned before, the main
difference between neuronal modeling and connectionist logic systems is in their biological plausibility. All
systems in the first group implement particular neuron models and aim to accurately reproduce the low-level
brain processes. The systems in the second group are based on artificial neural networks. Despite being
heavily influenced by neuroscience and the ability to closely model certain elements of human cognition, the
biological plausibility of these models is not claimed by their authors.
In conclusion, as can be seen in Figure 3, hybrid architectures are the most numerous and diverse group,
showing the tendency to grow even more, thus confirming a prediction made almost a decade ago [128]. Hybrid
architectures form a continuum between emergent and symbolic systems depending on the proportions and
roles played by symbolic and sub-symbolic components. Although quantitative analysis of this space is not
feasible, it is possible to crudely subdivide it. For instance, some architectures such as CogPrime and Sigma
are conceptually closer to emergent systems as they share many properties with the neural networks. On
the other hand, REM, CHREST, and RALPH, as well as the architectures implementing symbolic sub-
processing, e.g. 3T and ATLANTIS, are very much within the cognitivist paradigm. These architectures are
primarily symbolic but may utilize probabilistic reasoning and learning mechanisms.

4 Perception

Regardless of its design and purpose, an intelligent system cannot exist in isolation and requires an input
to produce any behavior. Although historically the major cognitive architectures focused on higher-level
reasoning, it is evident that perception and action play an important role in human cognition [15].
Perception can be defined as a process that transforms raw input into the system’s internal representation
for carrying out cognitive tasks. Depending on the origin and properties of the incoming data, multiple
sensory modalities are distinguished. For instance, five most common ones are vision, hearing, smell, touch
and taste. Other human senses include proprioception, thermoception, nocioception, sense of time, etc.
Naturally, cognitive architectures implement some of these as well as other modalities that do not have a
correlate among human senses such as symbolic input (using a keyboard or graphical user interface (GUI))
and various sensors such as LiDAR, laser, IR, etc.
Depending on its cognitive capabilities, an intelligent system can take various amounts and types of data
as perceptual input. Thus, this section will investigate the diversity of the data inputs used in cognitive
architectures, what information is extracted from these sources and how it is applied. The visualization
in Figure 4 addresses the first part of the question by mapping the cognitive architectures to the sensory
modalities they implement: vision (V), hearing (A), touch (T), smell (S), proprioception (P) as well as the
data input (D), other sensors (O) and multi-modal (M)10 . Note that the data presented in the diagram is
aggregated from many publications and most architectures do not demonstrate all listed sensory modalities
in a single model.
10 The taste modality is omitted as it is only featured in a single architecture, Brain-Based Devices (BBD), where it is simulated

by measuring the conductivity of the object [298].

12 Iuliia Kotseruba, John K. Tsotsos

ARS/SiMA
GMU

COG

CELTS

GLAIR
ML

-BIC
MID

C
D AC
N ET

DIAR
EC
Mi

Y
3T

IG
CA
A
cro

MD
OG
RA

S
M

OD
RC
P
LP

si

PR
S

e
D
H

em
SA

BB
CH

S
Si T

ch
L
SP

PR
AR gm

ys
i on

A
IS a A

l
pt

Po
Co M A C
JA A m
O IS u
CK bs g
D iP P Su Co
Po RA o bo
g am D R
ut E
AD V CS EX
AP RT
T CO
The SE
o SA
CAPS r
S ao M
ANIU

modalities
Copyca
t/Metac A-CR
at C R
E
Darw, Neurod
ynamics ART
HTM MACSi
IMPRINT Kismet

MAMID iCub

MAX CoSy
ICARU
iCog S
Mus CAR
S
NAR ACa
S
R FOR
A R
OM EP
R IC
CA
OS ST
AIS
CA M Ym
R- i
RE . AS r
h
rc AP MO
.A
n

d EX
to

en
Te

LI mpa
gy

m D
pa

m
r

Co

o A
me

c
Xa

Re
T-R

CH ANT
Ho

DIA

AT

R
AC

ni
No
DIS

ES
L

on
Co
R
CA

vam
le

MID

s
STA
AR

DUA
Leabra

gPr
ERE

BECC
Discip
AR

Shruti

DSO
CLARION

IS
en
AS

ime
L

te

not implemented
A

simulation
simulation/physical sensors
physical sensors
multi-modal

Fig. 4 A diagram showing sensory modalities of cognitive architectures. Radial segments correspond to cognitive architectures
and each track corresponds to a modality. Modalities are ordered based on their importance (calculated as a number of
architectures that implement this modality): vision (V), symbolic input (D), proprioception (P), other sensors (O), audition
(A), touch (T), smell (S) and multi-modal (M). Modalities plotted closer to the center of the diagram are more common.
The following coloring convention was used for the individual segments: white (the modality is not implemented), yellow (the
modality is simulated), orange (sensory modality demonstrated both in real and simulated domains) and red (physical sensors
used). Architecture labels are color-coded as follows: green - symbolic, blue - hybrid, red - emergent. Architectures are ordered
clockwise from 12 o’clock based on how many different sensory modalities they implement and to what extent (scores used for
ranking range from 0 when the modality is not implemented to 4 for multi-modal perception).
A Review of 40 Years in Cognitive Architecture Research Core Cognitive Abilities and Practical Applications 13

Several observations can be made from the visualization in Figure 4. For example, vision is the most
commonly implemented sense, however, more than half of the architectures use simulation for visual input
instead of the physical camera. Modalities such as touch and proprioception are mainly used in the physically
embodied architectures. Some senses remain relatively unexplored, for example, sense of smell is only featured
in three architectures (GLAIR [485], DAC [357] and PRS [523]). Overall, symbolic architectures by design
have limited perceptual abilities and tend to use direct data input as the only source of information (see left
side of the diagram). On the other hand, hybrid and emergent architectures (located mainly in the right half
of the diagram) implement a broader variety of sensory modalities both with simulated and physical sensors.
However, regardless of its source, incoming sensory data is usually not usable in the raw form (except, maybe,
for symbolic input) and requires further processing. Below we discuss various approaches to the problem of
efficient and adequate perceptual processing in cognitive architectures.

4.1 Vision

For a long time, vision was viewed as the dominating sensory modality based on the available experimental
evidence [424]. While recent works suggest a more balanced view of human sensory experience [502], cognitive
architectures research remains fairly vision-centric, as it is the most studied and the most represented sensory
modality. Even though in robotics various non-visual sensors (e.g. sonars, ultrasonic distance sensors) and
proprioceptive sensors (e.g. gyroscopes, compasses) are used for solving visual tasks such as navigation,
obstacle avoidance and search, visual input accounts for more than half of all implemented input modalities.
According to Marr [347] visual processing is composed of three different stages: early, intermediate and
late. Early vision is data-driven and involves parallel processing of the visual scene and extracts simple ele-
ments, such as color, luminance, shape, motion, etc. Intermediate vision groups elements into regions, which
are then further processed in the late stage to recognize objects and assign meaning to them using available
knowledge. Although not mentioned by Marr, visual attention mechanisms, emotion and reward systems
also influence all stages of visual processing [548]. Thus, perception and cognition are tightly interconnected
starting throughout all levels of processing.
We base our analysis of visual processing in the cognitive architectures on image understanding stages
described in [547]. These stages include: 1) detection and grouping of intensity-location-time values (re-
sults in edges, regions, flow vectors); 2) further grouping of edges, regions, etc. (produces surfaces, volumes,
boundaries, depth information); 3) identification of objects and their motions; 4) building object-centered
representations for entities; 5) assigning labels to objects based on the task; 6) inference of spatiotemporal
relationships among entities11 . Here only stage 1 represents early vision, and all subsequent stages require
additional task or world knowledge. Already at stage 2, grouping of features may be facilitated by the view-
point information and knowledge of the specific objects being viewed. Finally, the later stages involve spatial
reasoning and operate on high-level representations abstracted from the results of early and intermediate
processing.
Note that in computer vision research many of these image understanding stages are implemented im-
plicitly with deep learning methods. Given the success of deep learning in many applications it is surprising
to see that very few cognitive architectures employ it. Some applications of deep learning to simple vision
tasks can be found in CogPrime [196], LIDA [340], SPA [499] and BECCA [439].
Diagrams in Figure 5 show stages of processing implemented for real and simulated vision. We assume that
the real vision systems only receive pixel-level input with no additional information (e.g. camera parameters,
11 The last stage forming consistent internal descriptions - from [547] is omitted. In cognitive architectures, such representations

would be distributed among various reasoning and memory modules, making proper evaluation infeasible.
14 Iuliia Kotseruba, John K. Tsotsos

s
n

n
i a be l s

i a be l s
io

io
j s cts

j s cts
l r ls

l r ls
a t la e

a t la e
at

at
sp e ct o d

sp e ct o d
o b e ct b j e

o b e ct b j e
el

el
j m

j m
ob to- s

ob to- s
j o

j o
o e

o e
o b e ct

o b e ct
pr tur

pr tur
a

a
fe

fe
3T ACT-R NO DETAILS: NOT IMPLEMENTED:
4D-RCS APEX
ACT-R CELTS ARS/SiMA CAPS
ADAPT CLARION ASMO AIS
ARCADIA CoJACK CHARISMA CHREST
ARDIS Companions CogPrime COGNET
ART EPIC DiPRA Copycat/Metacat
ATLANTIS ERE Disciple Darwinian Neurodynamics
BBD FORR GMU-BICA HTM
BECCA GLAIR MicroPsi IMPRINT
CARACaS Homer MLECOG MAMID
CERA-CRANIUM ICARUS MAX
CORTEX MIDAS MusiCog
CSE MIDCA NARS
CoSy Novamente OMAR
DAC Pogamut OSCAR
DIARC Sigma PRODIGY
DSO Soar PRS
DUAL R-CAST
b) simulated vision RALPH
ISAC
Kismet Recommendation Architecture
LIDA REM
Leabra Shruti
MACSi Teton
MDB Theo
PolyScheme Ymir
RoboCog Xapagy
SAL
SASE
SPA
STAR
Soar
Subsumption
iCub
a) real vision
Fig. 5 A diagram showing the stages of real and simulated visual processing implemented by the cognitive architectures. The
stages are ordered from early to late processing: 1) features, 2) proto-objects, 3) objects, 4) object models, 5) object labels, 6)
spatial relations. Different shades of blue are used to indicate processes that belong to early, intermediate and late vision. The
architectures with real and simulated vision are shown in left and right columns respectively. The order within each column
is alphabetical. In some architectures some type of visual process is implemented, however, the publications do not provide a
sufficient amount of technical detail for our analysis (”No details” list). Architectures in the ”Not implemented” list do not
implement vision explicitly, although some may use other sensors (e.g. sonars and bumpers) for visual tasks such as recognition
and navigation.

locations and features of objects, etc.). The images should be produced by a physical camera, but the
architecture does not need to be connected to a physical sensor (i.e. if the input is a dataset of images or
previously recorded video, it is still considered real vision). The simulated vision systems generally omit early
and intermediate vision and receive input in the form that is suitable for the later stages of visual processing
(e.g. symbolic descriptions for shape and color, object labels, coordinates, etc.). Technically, any architecture
that does not have native support for real vision or other perceptual modalities, may be extended with an
A Review of 40 Years in Cognitive Architecture Research Core Cognitive Abilities and Practical Applications 15

interface that connects it to sensors and pre-processes raw data into a more suitable format (e.g. Soar [377]
and ACT-R [541]).
The illustration in Figure 5 shows what image interpretation stages are implemented, but it does not
reflect the complexity and extent of such processing as there are no quantitative criteria for evaluation. In the
remainder of this section we will provide brief descriptions of the visual processing in various architectures.

4.2 Vision using physical sensors

The majority of the architectures implementing all stages of visual processing are physically embodied and
include robot control, biologically inspired and biomimetic architectures. The architectures in the first group
(3T, ATLANTIS and CARACaS) operate in realistic unstructured environments and approach vision as an
engineering problem. Early vision (step 1) usually involves edge detection and disparity estimation. These
features are then grouped (step 2) into blobs with similar features (color, depth, etc.), which are resolved
into candidate objects with centroid coordinates (step 3). Object models are learned off-line using machine
learning techniques (step 4) and can be used to categorize candidate objects (step 5). For example, in RCS
architecture the scene is reconstructed from stereo images and features of the traversable ground are learned
from LiDAR data and color histograms [234]. Similarly, the ATLANTIS robot architecture constructs a height
map from stereo images and uses it to identify obstacles [367]. The CARACaS architecture also computes
a map for hazard avoidance from stereo images and range sensors. The locations and types of obstacles are
placed on the map relative to the robot itself. Its spatial reasoning (step 6) is limited to determining the
distance to the obstacles or target locations for navigation [250].
Biologically inspired architectures also make use of computer vision algorithms and follow similar pro-
cessing steps. For instance, neural networks for object detection (RCS [3], DIARC [471], Kismet [68]), SIFT
features for object recognition (DIARC [470]), SURF features, AdaBoost learning and a mixture of Gaussians
for hand detection and tracking (iCub [98]), Kinect and LBP features in conjunction with SVM classifier to
find people and determine their age and gender (RoboCog and CORTEX [354,442]).
In these architectures vision is more intertwined with memory and control systems and some steps in
visual processing are explicitly related to the human visual system. One such example is saliency, which
models the ability to prioritize visual stimuli based on their features or relevance to the task. As such,
saliency is used to find regions of interest in the scene (Kismet [60], ARCADIA [72], DIARC [473], iCub
[324], STAR [293]). Ego-sphere, a structure found in some robotic architectures, mimics functions of the
hippocampus in the integration of sensory information with action, although not in a biologically plausible
way [413]. Essentially, ego-sphere forms a virtual dome surrounding the robot, onto which salient objects
and events are mapped. Various implementations of this concept are included in RCS [5], ISAC [269], iCub
[569] and MACSi [257].
The architectures in the third subgroup pursue biologically plausible vision. One of the most elaborate
examples is the Leabra vision system called LVis [405] based on the anatomy of the ventral pathway of the
brain. It models the primary visual cortex (V1), extrastriate areas (V2, V4) and the inferotemporal (IT)
cortex. Computation in these areas roughly corresponds to early and intermediate processing steps on the
diagram. LVis possesses other features of the human visual system, such as larger receptive fields of neurons
in higher levels of the hierarchy, reciprocal connections between the layers and recurrent inhibitory dynamics
that limit activity levels across layers [609]. The visual systems of Darwin VIII (BBD) [481], SPA (Spaun)
[134] and ART [203] are also modeled on the primate ventral visual pathway.
The SASE architecture does not replicate the human visual system as closely. Instead, it uses a hierarchical
neural network with localized connections, where each neuron gets input from a restricted region in the
16 Iuliia Kotseruba, John K. Tsotsos

previous layer. The sizes of receptive fields within one layer are the same and increase at higher levels [617].
This system was tested on a SAIL robot in an indoor navigation scenario [591]. A similar approach to vision
is implemented in MDB [131], BECCA [440] and DAC [357].
Note that although emergent systems do not explicitly assign labels to objects, they are capable of forming
representations of spatial relations between the objects in the scene and use these representations for visual
tasks like navigation (BBD [162], BECCA [440], DAC [357], MDB [131], SASE [588]).

4.3 Simulated vision

As evident from the diagram in Figure 5, most of the simulations support only the late stages of visual
processing. The simplest kind of simulation is a 2D grid of cells populated by objects, e.g. NASA TileWorld
[418] used by ERE [127] and PRS [284], Wumpus World for GLAIR agents [485], 2D maze for FORR
agent Ariadne [139] and tribal simulation designed for CLARION social agents [510]. An agent in grid-
like environments often has a limited view of the surroundings, restricted to few cells in each direction.
Blocks world is another classic domain, where the general task is to build stacks of blocks of various shapes
and colors (ACT-R [274], ICARUS [314], MIDCA [109]. Despite their varying complexity and purpose,
different simulations usually provide the same kinds of data about the environment: objects, their properties
(color, shape, label, etc.), locations and properties of the agent itself, spatial relations between objects and
environmental factors (e.g. weather and wind direction). Such simulations mainly serve as a visualization
tool and are not too far removed from the direct data input since little to none sensory processing is needed.
More advanced simulations represent the scene as polygons with color and 3D coordinates of corners,
which have to be further processed to identify objects (Novamente [221]). Otherwise, the visual realism of
3D simulations is mostly for aesthetics and sensory, as the information is available directly in symbolic form
(e.g. CoJACK [145], Pogamut [50]).
As mentioned earlier, the diagram in Figure 5 does not reflect the differences in the complexity of the
environments or capabilities of the individual architectures. However, there is a lot of variation in terms of
size and realism among environments for the embodied cognitive architectures. For example, a planetary
rover controlled by ATLANTIS performs cross-country navigation in an outdoor rocky terrain [358], the
salesman robot Gualzru (CORTEX [442]) moves around a large room full of people and iCub (MACsi [256])
recognizes and picks up various toys from a table. On the other hand, simple environments with no clutter
or obstacles are also used in the cognitive architectures research (BECCA [440], MDB [42]). In addition,
color-coding objects is a common way of simplifying visual processing. For instance, ADAPT tracks a red
ball rolling on the table [48] and DAC orients itself towards targets marked with different colors [341].
Furthermore, most reviewed visual systems can recognize only a handful of different object categories.
In our selection, only Leabra is able to distinguish dozens of object categories [609]. The quality of visual
processing is greatly improved with the spread of available software toolkits such as OpenCV, Cloud Point
Library or Kinect API. But not as much progress has been made within systems that try to model general
purpose and biologically plausible visual systems. Currently, their applications are limited to controlled
environments.

4.4 Audition

Audition is a fairly common modality in the cognitive architectures as sound or voice commands are typi-
cally used to guide an intelligent system or to communicate with it. Since the auditory modality is purely
A Review of 40 Years in Cognitive Architecture Research Core Cognitive Abilities and Practical Applications 17

functional, many architectures resort to using available speech-to-text software rather than develop models
of audition. Among the few architectures modeling auditory perception are ART, ACT-R, SPA and EPIC.
For example, ARTWORD and ARTSTREAM were used to study phonemic integration [205] and source
segregation (cocktail party problem) [204] respectively. A model of music interpretation was developed with
ACT-R [96].
Using dedicated software for speech processing and communication helps to achieve a high degree of
complexity and realism. For example, in robotics applications it allows a salesman robot to have scripted
interaction with people in a crowded room (CORTEX [442]) or dialog about the objects in the scene in a
subset of English (CoSy [331]). A more advanced application involves using speech recognition for the task
of ordering books from the public library by phone (FORR [142]). Other systems using off-the-shelf speech
processing software include the PolyScheme [540] and ISAC [413].
In our sample of architectures, most effort is directed at natural language processing, i.e. linguistic and
semantic information carried by the speech (further discussed in Section 11.2), and much less attention is
paid to the emotional content, (e.g. loudness, speech rate, and intonation). Some attempts in this direction
are made in social robotics. For example, the social robot Kismet does not understand what is being said to
it but can determine approval, prohibition or soothing based on the prosodic contours of the speech [63]. The
Ymir architecture also has a prosody analyzer combined with a grammar-based speech recognizer that can
understand a limited vocabulary of 100 words [536]. Even the sound itself can be used as a cue, for example,
the BBD robots can orient themselves toward the source of a loud sound [481].

4.5 Symbolic input

The symbolic input category in Figure 4 combines several input methods which do not fall under physical
sensors or simulations. These include input in the form of text commands and data, and via a GUI. Text input
is typical for the architectures performing planning and logical inference tasks (e.g. NARS [580], OSCAR
[421], MAX [299], Homer [566]). Text commands are usually written in terms of primitive predicates used in
the architecture, so no additional parsing is required.
Although many architectures have tools for visualization of results and the intermediate stages of compu-
tation, interactive GUIs are less common. They are mainly used in human performance research to simplify
input of the expert knowledge and to allow multiple runs of the software with different parameters (IMPRINT
[371], MAMID [248], OMAR [119], R-CAST [178]).
The data input can be in text or any other format (e.g. binary or floating point arrays) and is primarily
used for the categorization and classification applications (e.g. HTM [320], CSE [223], ART [89]).

4.6 Other modalities

Other sensory modalities such as touch, smell and proprioception are trivially implemented with a variety
of sensors (physical and simulated). For instance, many robotic platforms are equipped with touch-sensitive
bumpers for an emergency stop in case of hitting an obstacle (e.g. RCS [7], AIS [216], CERA-CRANIUM
[20], DIARC [545]). Similarly, proprioception, i.e. sensing of the relative positions of the body parts and
effort being exerted, can be either simulated or provided by force sensors (3T [159], iCub [410], MDB [45],
Subsumption [77]), joint feedback (SASE [236], RCS [58]), accelerometers (Kismet [64], 3T [608]), etc.
18 Iuliia Kotseruba, John K. Tsotsos

4.7 Multi-modal perception

In the previous sections we have considered all sensory modalities in isolation. In reality, however, the human
brain receives a constant stream of information from different senses and integrates it into a coherent world
representation. This is also true for cognitive architectures, as nearly half of them have more than two
different sensory modalities (see Figure 4). As mentioned in the beginning of this section, not all of these
modalities may be present in a single model and most architectures utilize only two different modalities
simultaneously, e.g. vision and audition, vision and symbolic input or vision and range sensors. With few
exceptions, these architectures are embodied and essentially perform what is known as feature integration in
cognitive science [620] or sensor data fusion in robotics [276]. Evidently, it is possible to use different sensors
without explicitly combining their output, as demonstrated by the Subsumption architecture [164], but this
approach is rather uncommon.
Multiple sensory modalities improve the robustness of perception through complementarity and redun-
dancy but in practice using many different sensors introduces a number of challenges, such as incom-
plete/spurious/conflicting data, data with different properties (e.g. dimensionality or value ranges), the need
for data alignment and association, etc. These practical issues are well researched by the robotics commu-
nity, however, no universal solutions have been proposed. Eventually, every solution must be custom-made
for a particular application, a common approach taken by the majority of cognitive architectures. There
is, unfortunately, little technical information in the literature to determine the exact techniques used and
connect them to the established taxonomies (e.g. from a recent survey [276]).
Overall, any particular implementation of sensory integration depends on the representations used for
reasoning and the task. In typical robotic architectures with symbolic reasoning, data from various sensors
is processed independently and mapped onto a 3D map centered on the agent that can be used for nav-
igation (CaRACAS [135], CoSy [97]). In social robotics applications, as we already mentioned, the world
representation can take the form of an ego-sphere around the agent that contains ego-centric coordinates
and properties of the visually detected objects, which are associated to the location of the sound determined
via triangulation (ISAC [412], MACsi [19]). In RCS, a model with a hierarchical structure, there is a sensory
processing module at every level of the hierarchy with a corresponding world representation (e.g. pixel maps,
3D models, state tables, etc.) [477].
Some architectures do not perform data association and alignment explicitly. Instead, sensor data and
feature extraction (e.g. coordinates of objects from the camera and distances to the obstacles from laser)
are done independently and concurrently. The extracted information is then directly added to the working
memory, the blackboard or an equivalent structure (see Section 7.2 for more details). Any ambiguities and
inconsistencies are resolved by higher-order reasoning processes. This is a common approach in distributed
architectures, where independent modules concurrently work towards achieving a common goal (e.g. CERA-
CRANIUM [20], Polyscheme [95], RoboCog [80], Ymir [534] and LIDA [339]).
In many biologically inspired architectures the association between the readings of different sensors (and
actions) is learned. For example, DAC uses Hebbian learning to establish data alignment for mapping neural
representations of different sensory modalities to a common reference frame, mimicking the function of the
superior colliculus of the brain [356]. ART integrates visual and ultrasonic sensory information via neural
fusion (ARTMAP network) for mobile robot navigation [351]. Likewise, MDB uses neural networks to learn
the world model from sensor inputs and a genetic algorithm to adjust the parameters of the networks [45].
All approaches mentioned so far bear some resemblance to human sensory integration as they use the
spatial and temporal proximity and/or learning to combine and disambiguate multi-modal data. But over-
all, only few architectures aim for biological fidelity at the perceptual level. The only elaborate model of
biologically-plausible sensory integration is demonstrated using a brain-based device (BBD) architecture.
A Review of 40 Years in Cognitive Architecture Research Core Cognitive Abilities and Practical Applications 19

The embodied neural model called Darwin XI is constructed to investigate the integration of multi-sensory
information (from touch sensors, laser, camera and magnetic compass) and the formation of place activity in
the hippocampus during maze navigation [163]. The neural network of Darwin XI consists of approximately
80,000 neurons with 1.2 million synapses and simulates 50 neural areas. In a lesion study, the robustness
of the system was demonstrated by removing one or several sensory inputs and remapping of the sensory
neuronal units.
Generally, cognitive architectures largely ignore cross-modal interaction. The architectures, including
biologically- and cognitively-oriented ones, typicaly pursue a modular approach when dealing with different
sensory modalities. At the same time, many psychological and neuroimaging experiments conducted in the
past decades suggest that sensory modalities mutually affect one another. For example, vision alters auditory
processing (e.g. McGurk effect [362]) and vice versa [488]. While, some biomimetic architectures, such as
BBD mentioned above, may be capable of representing cross-modal effects, to the best of our knowledge,
this problem has not been investigated yet.

5 Attention

Perceptual attention plays an important role in human cognition, as it mediates the selection of relevant and
filters out irrelevant information from the incoming sensory data. However, it would be wrong to think of
attention as a monolithic structure that makes a decision about what to process next. Rather, the opposite
may be true. There is ample evidence in favor of attention as set of mechanisms affecting both perceptual
and cognitive processes [551]. Currently, visual attention remains the most studied form of attention, as
there are no comprehensive frameworks for other sensory modalities. Since only a few architectures have
rudimentary mechanisms for modulating auditory data (OMAR [122], iCub [449], EPIC [282] and MIDAS
[199]), this section will be dedicated to visual attention.
For the following analysis, we use the taxonomy of attentional mechanisms proposed by Tsotsos [548].
Elements of attention are grouped into three classes of information reduction mechanisms: selection (choose
one from many), restriction (choose some from many) and suppression (suppress some from many). Selection
mechanisms include gaze and viewpoint selection, world model (selection of objects/events to focus on) and
time/region/features/objects/events of interest. Restriction mechanisms serve to prune the search space by
priming12 (preparing the visual system for input based on task demands), endogenous motivations (domain
knowledge), exogenous cues (external stimuli), exogenous task (restrict attention to objects relevant for the
task), and visual field (limited field of view). Suppression mechanisms consist of feature/spatial surround
inhibition (temporary suppression of the features around the object while attending), task irrelevant stimuli
suppression, negative priming, and location/object inhibition of return (a bias against returning attention
to previous attended location or stimuli). Finally, branch-and-bound mechanisms combine elements of sup-
pression, selection and restriction. For more detailed explanations of the attentional mechanisms and review
of the relevant psychological and neuroscience literature refer to [548].
The diagram in Figure 6 shows a summary of the implemented visual attention elements in the cognitive
architectures. Here we only consider the architectures with implemented real or simulated vision and omit
projects for which not enough technical details are provided (same as in the previous section). It is evident
that most of the implemented mechanisms of attention belong to the selection and restriction group.
Only a handful of architectures implement suppression mechanisms. For instance, suppression of task
irrelevant visual information is done in only three architectures: in BECCA irrelevant features are suppressed
12 Here we consider only priming in vision, other types of priming and their importance for learning are discussed in Section
8.6
20 Iuliia Kotseruba, John K. Tsotsos

ns

ns
or o i j n g n d

or o i j n g n d
vi ge ou s m rest

v i e u m re s t
io

io
ou

ou
an fie s t es at

an fie s t es at
re e o od aze IOR

re e o od aze IOR
te

te
nd

nd
vi atio e p ant urr

vi atio e p ant urr

of s t

br al ou cu tiv

of s t

br al ou cu tiv
in

in
en min eve inte st
d g n t re

en min eve inte st

d g n t re
su n s o

ou

su n s o

ou
s

s
k

k
pr ject s o tere

pr ject s o tere
t

e n t

ch ld as

m g t

e n t

ch ld as
lo at elev tial

w p /ob mi

lo at elev tial

w wp /ob imi
tim ld nt/ ec

ob ur f i e s

tim ld nt/ ec

ob tur of i res
-b

-b
og no us
fe ion inte l

fe ion inte l
ew n ri

g f e

g f e
at o r

nd

nd
g

o n u
g r a

g r a

e n r
i / f

i / f
ex ge no

ex ge no
ne k ir /sp

ne k ir /sp
-a

-a
m

o e

o e
c iv

c iv
s e

s e
ex og

ex og
ta tur

ta tur
a

a
fe

fe
3T ACT-R
4D-RCS APEX
ACT-R CELTS
ADAPT CLARION
ARCADIA CoJACK
ARDIS Companions
ART EPIC
ATLANTIS ERE
BBD FORR
BECCA GLAIR
CARACaS Homer
CERA-CRANIUM ICARUS
CORTEX MIDAS
CSE MIDCA
CoSy Novamente
DAC Pogamut
DIARC Sigma
DSO Soar
DUAL Suppress Select Restrict
ISAC
Kismet b) simulated vision
LIDA
Leabra
MACSi
MDB
PolyScheme
RoboCog
SAL
SASE
SPA
STAR
Soar
Subsumption
iCub
Suppress Select Restrict
a) real vision

Fig. 6 A diagram showing visual attention mechanisms implemented in cognitive architectures for suppressing, selecting and
restricting visual information. Architectures are ordered alphabetically. As in Section 4.1, architectures in the left and right
columns implement real and simulated vision respectively. Architectures not implementing vision or missing technical details
are not shown. Suppression, selection and restriction mechanisms in the diagram are shown in shades of blue, yellow-red and
green respectively.

through the WTA mechanisms [437], in DSO background features are suppressed to speed up processing
[610], in MACSi depth is used to ignore regions that are not reachable by the robot [337] and in ARCADIA
non-central regions of the visual field are inhibited at each cycle since the cue always appears in the center
[72]. Another suppression mechanism is inhibition of return (IOR), which prevents the visual system from
attending to the same salient stimuli. In ACT-R activation values and distance are used to ignore objects
for consecutive WHERE requests [397]. In ART, iCub and STAR an additional map is used to keep records
of the attended locations. The temporal nature of inhibition can be modeled with a time decay function, as
it is done in iCub [449]. ARCADIA mentions a covert inhibition mechanism, however, the implementation
details are not provided [72].
Selection mechanisms are more commonly found in the cognitive architectures. For example, a world
model by default is a part of any visual system. Viewpoint/gaze selection is a necessary component of active
A Review of 40 Years in Cognitive Architecture Research Core Cognitive Abilities and Practical Applications 21

vision systems. Gaze control allows to focus on a region in the environment and viewpoint selection allows to
get more information about the region/object by changing the distance to it or by viewing it from different
angles. The physically embodied architectures automatically support viewpoint selection because a camera
installed on a robot can be moved around the environment. Eye movements are usually simulated, but it is
also possible to implement them on a humanoid robot (iCub, Kismet).
Other mechanisms in this group include selection of various features/objects of interest. The selection of
visual data to attend can be data-driven (bottom-up) or task-driven (top-down). The bottom-up attentional
mechanisms identify salient regions whose visual features are distinct from the surrounding image features,
usually along a combination of dimensions, such as color channels, edges, motion, etc. Some architectures
resort to the classical visual saliency algorithms, such as Guided Search [607] used in ACT-R [397] and
Kismet [67], the Itti-Koch-Niebur model [255] used by ARCADIA [72], iCub [449] and DAC [356] or AIM
[79] used in STAR [293]. Other approaches include filtering (DSO [610]), finding unusual motion patterns
(MACsi [257]) or discrepancies between observed and expected data (RCS [10]).
Top-down selection can be applied to further limit the sensory data provided by the bottom-up process-
ing. For example, in visual search, knowing desired features of the object (e.g., the color red) can narrow
down the options provided by the data-driven figure-ground segmentation. Many architectures resort to this
mechanism to improve search efficiency (ACT-R [455], APEX [174], ARCADIA [46], CERA-CRANIUM [20],
CHARISMA [100], DAC [356]). Another option is to use a hard-coded or learned heuristics. For example,
CHREST looks at typical positions on a chess board [312] and MIDAS replicates common eye scan patterns
of pilots [199]. The limitation of the current top-down approaches is that they can direct vision for only a
limited set of predefined visual tasks, however, ongoing research in STAR attempts to address this problem
[551,548].
Restriction mechanisms allow to further reduce the complexity of vision by limiting the search space
to certain features (priming), events (exogenous cues), space (visual field), objects (exogenous task) and
knowledge (endogenous motivations). Exogenous task and endogenous motivations are implemented in all
the architectures by default as domain knowledge and task instructions. In addition to that, attention can
be modulated by internal signals such as motivation, emotions, which are discussed in the next section on
action selection.
A limited visual field is a feature of any physically embodied system since most cameras do not have a
360-degree field of view. This feature is also present in some simulated vision systems. An ability to react
to the sudden events (exogenous cues), such as fast movement or bright color, is common in social robotics
applications (e.g. ISAC [270], Kismet [65]). On the other hand, priming allows to bias the visual system
towards particular types of stimuli via task instruction. For example, by assigning more weight to skin-
colored features during the saliency computation to improve human detection (Kismet [60]) or by utilizing
the knowledge of the probable location of some objects to spatially bias detection (STAR [293]). Neural
mechanisms of top-down priming are studied in ART models [88].
Overall, visual attention is largely overlooked in cognitive architectures research with the exception of
the biologically plausible visual models (e.g. ART) and the architectures that specifically focus on vision
research (ARCADIA, STAR13 ). This is surprising because strong theoretical arguments as to its importance
in dealing with the computational complexity of visual processing have been known for decades [546]. Often
the attention mechanisms found in the cognitive architectures are side-effects of other design decisions.
For example, task constraints, world model and domain knowledge are necessary for the functioning of other
13 Note that biologically plausible models of vision, such as Selective Tuning [550], extensions to HMAX [573, 574, 432] and

others [177], support many of the listed attention mechanisms. However, embedding these models in the cognitive architecture
is non-trivial. For instance, STAR represents ongoing work on integrating general purpose vision, as represented by the Selective
Tuning model, with other higher-order cognitive processes.
22 Iuliia Kotseruba, John K. Tsotsos

aspects of intelligent system and are implemented by default. Limited visual field and viewpoint changes often
result from physical embodiment. Otherwise, region of interest selection and visual reaction to exogenous
cues are the two most common mechanisms explicitly included for optimizing visual processing.
In the cognitive and psychological literature, attention is also used as a broad term for the allocation of
limited resources [444]. For instance, in the Global Workspace Theory (GWT) [28] attentional mechanisms
are central to perception, cognition and action. According to GWT, the nervous system is organized as
multiple specialized processes running in parallel. Coalitions of these processes compete for attention in the
global workspace and the contents of the winning coalition are broadcast to all other processes. For instance,
the LIDA architecture is an implementation of GWT14 [172]. Other architectures influenced by the GWT
include ARCADIA [72] and CERA-CRANIUM [21].
Along the same lines, cognition is simulated as a set of autonomous independent processes in ASMO
(inspired by Society of Mind [369]). Here an attention value for each process is either assigned directly or
learned from experience. Attention values vary dynamically and affect action selection and resource allocation
[392]. A similar idea is implemented in COGNET [616], DUAL [287] and Copycat/Metacat [373] in which
multiple processes also compete for attention. Other computational mechanisms, such as queue (PolyScheme
[301]), Hopfield nets (CogPrime [254]) and modulators (MicroPsi [32]) have been used to implement a focus
of attention. In MLECOG, in addition to saccades that operate on visual data, mental saccades allow to
switch between the most activated memory neurons [497].

6 Action selection

Informally speaking, action selection determines at any point in time ”what to do next”. At the highest
level, action selection can be split into the what part involving the decision making and the how part
related to motor control [406]. However, this distinction is not always explicitly made in the literature,
where action selection may refer to goal, task or motor command. For example, in the MIDAS architecture
action selection involves both the selection of the next goal to pursue and of actions that implement it
[553]. Similarly, in MIDCA the next action is normally selected from a planned sequence if one exists. In
addition to this, a different mechanism is responsible for the adoption of the new goals based on dynamically
determined priorities [408]. In COGNET and DIARC, selection of a task/goal triggers an execution of the
associated procedural knowledge [614,71]. In DSO, the selector module chooses between available actions
or decisions to reach current goals and sub-goals [388]. Since the treatment of action selection in various
cognitive architectures is inconsistent, in the following discussion the action selection mechanisms may apply
both to decision-making and motor actions.
The diagram in Figure 7 illustrates all implemented action selection mechanisms organized by the type
of the corresponding architecture (symbolic, hybrid and emergent). We distinguish between two major ap-
proaches to action selection: planning and dynamic action selection. Planning refers to the traditional AI
algorithms for determining a sequence of steps to reach a certain goal or to solve a problem in advance
of their execution. In dynamic action selection, one best action is chosen among the alternatives based on
the knowledge available at the time. For this category, we consider the type of selection (winner-take-all
(WTA), probabilistic, predefined) and criteria for selection (relevance, utility, emotion). The default choice
is always the best action based on the defined criteria (e.g. action with the highest activation value). Reactive
actions are executed bypassing action selection. Finally, learning can also affect action selection but will be
discussed in Section 8. Note that these action selection mechanisms are not mutually exclusive and most
14 In GWT focus of attention is associated with consciousness, however, this topic is beyond the scope of this survey.
A Review of 40 Years in Cognitive Architecture Research Core Cognitive Abilities and Practical Applications 23

s
r

r
e cto

e cto

e cto
ut levafine tic

ut levafine tic

ut levafine tic
in lity nced

in lity nced

in lity nced
tiv fa

tiv fa

tiv fa
re e d e b i l i s

re e d e b i l i s

re e d e b i l i s
pr TA ing

pr TA ing

pr TA ing
ac al

ac al

ac al
p ro b a

p ro b a

p ro b a
re ern

re ern

re ern
W nn

W nn

W nn
a

a
i
t

i
t

i
t
pl

pl

pl
ERE CARACaS BECCA
MusiCog RALPH MicroPsi
Companions Pogamut BBD
MAX Soar Darw. Neurodyn.
PRODIGY ADAPT DAC
Disciple CHARISMA Leabra
GLAIR DIARC SPA
R-CAST DiPRA ART
MIDAS CoJACK Shruti
ICARUS MIDCA HTM
PRS RoboCog Subsumption
APEX ARS/SiMA Recommendation
Homer 3T SASE
Theo ATLANTIS MDB
IMPRINT CORTEX
OSCAR CoSy c) emergent
Teton RCS
OMAR ARDIS
COGNET REM
EPIC iCub
AIS CELTS
MAMID LIDA
Copycat/Metacat
a) symbolic FORR
ACT-R
CHREST
CLARION
CogPrime
Novamente
CSE
Sigma
Ymir
ARCADIA
Polyscheme
DUAL
GMU-BICA
STAR
CAPS
SAL
Xapagy
ASMO
CERA-CRANIUM
ISAC
MACSi
NARS
MLECOG
DSO
Kismet

b) hybrid
Fig. 7 A diagram of mechanisms involved in action selection. The visualization is organized in three columns grouping symbolic,
hybrid and emergent architectures. Note that in this diagram (as well as in the diagrams for the sections 7 and 8) the sorting order
emphasizes clusters of architectures with similar action selection mechanisms (or memory and learning approaches respectively).
Since the discussion follows the order in which different mechanisms are shown in the diagram, it can help to identify groups of
architectures focusing on a particular mechanism which correspond to clusters within each column. A version of the diagram
with an alphabetical ordering of the architectures is also available on the project website.
24 Iuliia Kotseruba, John K. Tsotsos

of the architectures have more than one. Even though few architectures implement the same set of action
selection mechanisms (as can be seen in the Figure 7), the whole space of valid combinations is likely much
larger. Below we discuss approaches to action selection and what cognitive processes they correspond to in
more detail.

6.1 Planning vs reactive actions

Predictably, planning is more common in the symbolic architectures, but can also be found in some hybrid and
even emergent (MicroPsi [32]) architectures. In particular, task decomposition, where the goal is recursively
decomposed into subgoals, is a very common form of planning (e.g. Soar [117], Teton [560], PRODIGY
[158]). Other types of planning are also used: temporal (Homer [567]), continual (CoSy [210]), hierarchical
task network (PRS [124]), generative (REM [381]), search-based (Theo [375]), hill-climbing (MicroPsi [32]),
etc
Very few systems in our selection rely on classical planning alone, namely OSCAR, used for logical
inference, and IMPRINT, which employs task decomposition for modeling human performance. Otherwise,
planning is usually augmented with more dynamic action selection mechanisms to improve adaptability to
the changing environment.
On the other end of the spectrum are reactive actions. These are executed immediately, suspending any
ongoing activity and bypassing reasoning, similar to reflex actions in humans as automatic responses to
stimuli. As demonstrated by the Subsumption architecture, combining multiple reactive actions can give rise
to complex behaviors [77], however, purely reactive systems are rare and reactive actions are used only under
certain conditions. For example, they are used to protect the robot from the collision (ATLANTIS [179], 3T
[56]) or to automatically respond to unexpected stimuli, such as fast moving objects or loud sounds (ISAC
[269], Kismet [66], iCub [449]).

6.2 Dynamic action selection

Dynamic action selection offers the most flexibility and can be used to model many phenomena typical for
human and animal behavior.
Winner-take-all is a selection process in neuronal networks where the strongest set of inputs is amplified,
while the output from others is inhibited. It is believed to be a part of cortical processes and is often used in
the computational models of the brain to make a selection from a set of decisions depending on the input.
WTA and its variants are common in the emergent architectures such as HTM [83], ART [90], SPA [133],
Leabra [402], DAC [570], Darwinian Neurodynamics [156] and Shruti [584]. Similar mechanisms are also used
to find the most appropriate action in the architectures, where behavior emerges as a result of cooperation
and competition of multiple processes running in parallel (e.g. Copycat/Metacat [348], CELTS [149], LIDA
[171]).
A predefined order of action selection may serve different purposes. For example, in the Subsumption
architecture robot behavior is represented by a hierarchy of sub-behaviors, where higher-level behaviors over-
ride (subsume) the output of lower-level behaviors [77]. In FORR, the decision-making component considers
options from advisors in the order of increasing expertise to achieve robust and human-like learning [141].
In Ymir priority is given to processes within the reactive layer first, followed by the content layer and then
by the process control layer. Here the purpose is to provide a smooth real-time behavior generation. Each
layer has a different upper bound on the perception-action time, thus reactive modules provide automatic
A Review of 40 Years in Cognitive Architecture Research Core Cognitive Abilities and Practical Applications 25

feedback to the user (changing facial expressions, automatic utterances) while deliberative modules generate
more complex behaviors [535].
The remaining action selection mechanisms include finite-state machines (FSM), which are frequently
used to represent sequences of motor actions (ATLANTIS [179], CARACaS [250], Subsumption [77]) and even
to encode the entire behavior of the system (ARDIS [353], STAR [293]). Probabilistic action selection
is also common (iCub [569], CSE [223], CogPrime [190], Sigma [556], Darwinian Neurodynamics [156], ERE
[272], Novamente [193]).

6.2.1 Action selection criteria

There are several criteria that can be taken into account when selecting the next action: relevance, utility
and affect (which includes motivations, affective states, emotions, moods, drives, etc.).
Relevance reflects how well the action corresponds to the current situation. This mainly applies to
systems with symbolic reasoning and involves checking pre- and/or post-conditions of the rule before applying
it (MAX [300], Disciple [529], EPIC [278], GLAIR [484], PRODIGY [565], MIDAS [553], R-CAST [153],
Disciple [529], Companions [230], Ymir [535], Pogamut [74], Soar [307], ACT-R [321]).
Utility of the action is a measure of its expected contribution to achieving the current goal (CERA-
CRANIUM [21], CHARISMA [100], DIARC [71], MACSi [257], MAMID [242], NARS [581], Novamente [193]).
Some architectures also perform a ”dry run” of candidate actions and observe their effect to determine their
utility (MLECOG [497], RoboCog [37]). Utility can also take into account the performance of the action
in the past and improve the behavior in the future via reinforcement learning (ASMO [391], BECCA [435],
CLARION [518], PRODIGY [85], CSE [223], CoJACK [145], DiPRA [415], Disciple [529], DSO [388], FORR
[141], ICARUS [482], ISAC [269], MDB [41], MicroPsi [31], Soar [307]). Other machine learning techniques
are also used to associate goals with successful behaviors in the past (MIDCA [408], SASE [587], CogPrime
[190]).
Finally, internal factors do not determine the next behavior directly, but rather bias the selection.
For simplicity, we will consider short-term, long-term and lifelong factors corresponding to emotions, drives
and personality traits in humans. Given the impact of these factors on human decision making and other
cognitive abilities it is important to model emotion and affect in cognitive architectures, especially in the
areas of human-computer interaction, social robotics and virtual agents. Not surprisingly, this is where most
of the effort has been spent so far (see Section 11.2.9 on applications of intelligent agents with realistic
personalities and emotions).
Artificial emotions in cognitive architectures are typically modeled as transient states (associated
with anger, fear, joy, etc.) that influence cognitive abilities [240,195,598]. For instance, in CoJACK morale
and fear emotions can modify plan selection. As a result, plans that confront a threat have higher utility
when morale is high, but lower utility when fear is high [145]. Other examples include models of stress
affecting decision-making (MAMID [247]), emotions of joy/sadness affecting blackjack strategy (CHREST
[476]), analogical reasoning in the state of anxiety (DUAL [157]), effect of arousal on memory (ACT-R [99]),
positive and negative affect depending on goal satisfaction (DIARC [474]) and emotional learning in HCI
scenario (CELTS [149]). The largest attempt at implementing a theory of appraisal led to the creation of
Soar-Emote, a model based on the ideas of Damasio and Gratch & Marsella [345].
Drives are another source of internal motivations. Broadly speaking, they represent the basic physio-
logical needs such as food and safety, but can also include high-level or social drives. Typically, any drive
has strength that diminishes as the drive is satisfied. As the agent works towards its goals, several drives
may affect its behavior simultaneously. For example, in ASMO three relatively simple drives, liking color
red, praise and happiness of the user and the robot, bias action selection by modifying the weights of the
26 Iuliia Kotseruba, John K. Tsotsos

corresponding modules [391]. In CHARISMA preservation drives (avoid harm and starvation) combined with
curiosity and desire for self-improvement guide behavior generation [101]. In MACSi curiosity drives explo-
ration towards areas where the agent learns the fastest [389]. Likewise, in CERA-CRANIUM curiosity, fear
and anger affect exploration of the environment by the mobile robot [24]. The behavior of social robot Kismet
is organized around satiating three homeostatic drives: to engage people, to engage toys, and to rest. These
drives and external events, contribute to the affective state (or ”mood”) of the robot and its expression as
anger, disgust, fear, joy, sorrow and surprise via facial gestures, stance or changes in the tone of its voice.
Although its motivational mechanisms are hard-coded, Kismet demonstrates one of the widest repertoire of
emotional behaviors of all the cognitive architectures we reviewed [61].
Unlike emotions, which have a transient nature, personality traits are unique long-term patterns of be-
havior that manifest themselves as consistent preferences in terms of internal motivations, emotions, decision-
making, etc. Most of the identified personality traits can be reduced to a small set of dimensions/factors
necessary and sufficient for broadly describing human personality (e.g. the Five-Factor Model (FFM) [361]
and Mayers-Briggs model [384]). Likewise, personality in cognitive architectures is often represented by sev-
eral factors/dimensions, not necessarily based on a known model. These traits, in turn, are related to the
emotions and/or drives that the system is likely to experience.
In the simplest case a single parameter suffices to create a systematic bias in the behavior of the system.
Both NARS [577] and CogPrime [196] use the ”personality parameter” to define how much evidence is
required to evaluate the truth of the logical statement or to plan the next action respectively. The larger the
value of the parameter, the more ”conservative” the system appears to be.
In Novamente personality traits of a virtual animal (e.g. aggressiveness, curiosity, playfullness) are linked
to emotional states and actions via probabilistic rules [195]. Similarly, in AIS traits (e.g. nasty, phlegmatic,
shy, confident, lazy) are assigned an integer value that defines the degree to which they are exhibited.
Abstract rules define what behaviors are more likely given the personality [448]. In CLARION personality
type determines the baseline strength and initial deficit (i.e. proclivity towards) of the many predefined
drives. The mapping is embodied in a pretrained neural network [508].
Remarkably, even these simple models can generate a spectrum of personalities. For instance, the Pogamut
agent has 9 possible states and 5 personality factors (based on FFM) resulting in 45 distinct mappings. Each
of these mappings can produce any of the 12 predefined intentions to a varying degree and generate a wide
range of behaviors [446].
CLARION and MAMID deserve a special mention. CLARION provides a cognitively-plausible framework
that is capable of addressing emotions, drives and personality traits and relating them to other cognitive
processes including decision-making. Three aspects of emotion are modeled: reactive affect (or unconscious
experience of emotion), deliberative appraisal (potentially conscious) and coping/action (following appraisal).
Thus emotion emerges as an interaction between the explicit and implicit processes and involves (and affects)
perception, action and cognition [515,508]. Several CLARION models have been validated using psychological
data, computational model of the FFM personalities [514], coping with bullying at school [601], performance
degradation under pressure [602] and model of stereotype bias induced by social anxiety [603].
MAMID is a model of both generation and effect of emotions resulting from external events, internal
interpretations, goals and personality traits. Internally, belief nets relate the task- and individual-specific
criteria, e.g. whether a goal failure will result in anxiety in a particular agent [243]. MAMID has been
instantiated in two domains: peacekeeping mission training [238] and search-and-rescue tasks [241]. Of great
value are also contributions towards developing an interdisciplinary theory of emotions and design guidelines
[246,431].
A Review of 40 Years in Cognitive Architecture Research Core Cognitive Abilities and Practical Applications 27

7 Memory

Memory is an essential part of any systems-level cognitive model, regardless of whether the model is being
used for studying the human mind or for solving engineering problems. Thus, nearly all the architectures
featured in this review have memory systems that store intermediate results of computations, enabling learn-
ing and adaptation to the changing environment. However, despite their functional similarity, the particular
implementations of memory systems differ significantly and depend on the research goals and conceptual
limitations, such as biological plausibility and engineering factors (e.g. programming language, software ar-
chitecture, use of frameworks, etc.). In the cognitive architecture literature, memory is described in terms
of its duration (short- and long-term) and type (procedural, declarative, semantic, etc.), although it is not
necessarily implemented as separate knowledge stores.
The multi-store memory model is influenced by the Atkinson-Shiffrin model (1968) [27], later modified by
Baddeley [35]. This view of memory is dominant in psychology, but its utility for engineering is questioned
by some because it does not provide a functional description of various memory mechanisms [411]. Never-
theless, most architectures do distinguish between various memory types, although the naming conventions
differ depending on the conceptual background. For instance, the architectures designed for planning and
problem solving have short- and long-term memory storage systems but do not use terminology from cog-
nitive psychology. The long-term knowledge in planners is usually referred to as a knowledge base for facts
and problem-solving rules, which correspond to semantic and procedural long-term memory (e.g. Disciple
[524], MACsi [257], PRS [185], ARDIS [353], ATLANTIS [179], IMPRINT [70]). Some architectures also
save previously implemented tasks and solved problems, imitating episodic memory (REM [382], PRODIGY
[563]). The short-term storage in planners is usually represented by a current world model or the contents
of the goal stack.
Figure 8 shows a visualization of various types of memory implemented by the architectures. Here we
follow the convention of distinguishing between the long-term and short-term storage. Long-term storage is
further subdivided into semantic, procedural and episodic types, which store factual knowledge, information
on what actions should be taken under certain conditions respectively and episodes from the personal ex-
perience of the system respectively. Short-term storage is split into sensory and working memory following
[105]. Sensory or perceptual memory is a very short-term buffer that stores several recent percepts. Working
memory is a temporary storage for percepts that also contains other items related to the current task and is
frequently associated with the current focus of attention.

7.1 Sensory memory

The purpose of sensory memory is to cache the incoming sensory data and preprocess it before transferring
it to other memory structures. For example, iconic memory assists in solving continuity and maintenance
problems, i.e. identifying separate instances of the objects as being the same and retaining the impression
of the object when unattended (ARCADIA [72]). Similarly, echoic memory allows an acoustic stimulus to
persist long enough for perceptual binding and feature extraction, such as pitch extraction and grouping
(MusiCog [359]). The decay rate for items in sensory memory is believed to be tens of milliseconds (EPIC
[280], LIDA [29]) for visual data and longer for audio data (MusiCog [359]), although the time limit is not
always specified. Other architectures implementing this memory type include Soar [604], Sigma [443], ACT-R
[397], CHARISMA [102], CLARION [507], ICARUS [315] and Pogamut [74].
28 Iuliia Kotseruba, John K. Tsotsos

ob di al

ob di al

ob di al
epocentic

epocentic

epocentic
g l i so d u r

g l i so d u r

g l i so d u r
al c

al c

al c
seM ry

seM ry

seM ry
W n so

W n so

W n so
pr ma

pr ma

pr ma
se

se

se
ICARUS ACT-R BECCA
EPIC CHARISMA Leabra
MusiCog CLARION MicroPsi
APEX LIDA Recommendation
GLAIR Sigma SPA
Homer Soar Shruti
MAMID Pogamut SASE
OMAR ARCADIA MDB
PRODIGY ASMO ART
R-CAST CogPrime Darw. Neurodynamics
AIS NARS DAC
COGNET CELTS BBD
Disciple Copycat/Metacat HTM
ERE DUAL Subsumption
MAX GMU-BICA
MIDAS ISAC c) emergent
OSCAR MIDCA
PRS MLECOG
Teton Novamente
Theo REM
Companions Ymir
IMPRINT CORTEX
DiPRA
a) symbolic RALPH
RoboCog
3T
ADAPT
ARS/SiMA
ATLANTIS
CAPS
CERA-CRANIUM
CHREST
CoJACK
CoSy
DIARC
DSO
FORR
MACSi
PolyScheme
RCS
SAL
Xapagy
CARACaS
CSE
Kismet
STAR
ARDIS

b) hybrid
Fig. 8 A diagram showing types of memory implemented in the cognitive architectures. The symbolic, hybrid and emergent
architectures are shown in separate columns. The architectures where memory system is unified, i.e. no distinction is made
between short-, long-term and other types of memory, are labeled as ”global”. A version of the diagram with an alphabetical
ordering of the architectures is also available on the project website.
A Review of 40 Years in Cognitive Architecture Research Core Cognitive Abilities and Practical Applications 29

7.2 Working memory

Working memory can be defined as a mechanism for temporary storage of information related to the current
task. It is critical for cognitive capacities such as attention, reasoning and learning, thus every cognitive
architecture in our list implements it in some form. Particular realizations of working memory differ mainly in
what information is being stored, how it is represented, accessed and maintained. Furthermore, some cognitive
architectures contribute to ongoing research about the processes involved in encoding, manipulation and
maintenance of information in human working memory and its relationship to other processes in the human
brain. Despite the importance of working memory for human cognition, relatively few publications provide
sufficient detail about its internal organization and connections to other modules. Often it is summarized
in a few words, e.g. ”current world state” or ”data from the sensors”. Based on these statements we infer
that in many architectures working memory or an equivalent structure serves as a cache for the current
world model, the state of the system and/or current goals. Although there are no apparent limitations on
the capacity of working memory, new goals or new sensory data usually overwrite the existing content. This
simplified account of working memory can be found in many symbolic architectures (3T [55], ATLANTIS
[179], Homer [566], IMPRINT [70], MIDCA [110], PRODIGY [564], R-CAST [611], RALPH [398], REM
[381], Theo [375], Disciple [528], OSCAR [420], ARS/SiMA [468], CoJACK [434], ERE [69], MAX [299],
Companions [166], RCS [6]) as well as hybrids (CSE [223], MusiCog [359], MicroPsi [30], Pogamut [74]). More
cognitively plausible models of working memory use an activation mechanism. Some of the earliest models of
activation of working memory contents were implemented in ACT-R. As before, working memory holds the
most relevant knowledge, which is retrieved from the long-term memory as determined by bias. This bias is
referred to as activation and consists of base-level activation (which may decrease or increase) with every
access and may also include the spreading activation from neighboring elements. The higher the activation of
the element, the more likely it is to enter working memory and directly affect the behavior of the system [13].
This applies to the graph-based knowledge representation where nodes refer to concepts and weights assigned
to edges correspond to the associations between the concepts (Soar [304], CAPS [461], ADAPT [49], DUAL
[288], Sigma [426], CELTS [149], NARS [578], Novamente [192], MAMID [245]). Naturally, activation can be
used in neural network representation as well (Shruti [584], CogPrime [254], Recommendation Architecture
[106], SASE [590], Darwinian Neurodynamics [156]). Activation mechanisms contribute to modeling many
properties of working memory, such as limited capacity, temporal decay, rapid updating as the circumstances
change, connection to other memory components and decision-making. Another common paradigm, the
blackboard architecture, represents memory as a shared repository of goals, problems and partial results
which can be accessed and modified by the modules running in parallel (AIS [213], CERA-CRANIUM [23],
CoSy [490], FORR [137], Ymir [536], LIDA [171], ARCADIA [46], Copycat/Metacat [350], CHARISMA
[102], PolyScheme [94], PRS [183]). The solution to the problem is obtained by continuously updating the
shared short-term storage with information from specialized heterogeneous modules analogous to a group
of people completing a jigsaw puzzle. Although not directly biologically plausible, this paradigm works well
in situations where many disparate sources of information must be combined to solve a complex real-time
task. Finally, neuronal models of working memory based on the biology of the prefrontal cortex are realized
within SPA [500], ART [203] and Leabra [401]. They demonstrate a range of phenomena (validated on
human data) such as adaptive switching between rapid updating of the memory and maintaining previously
stored information [400] and reduced accuracy with time (e.g. in a list memory task) [500]. In some robotics
architectures, ego-sphere (discussed in Section 4) besides enabling spatiotemporal integration also provides
representation for the current location and the orientation of the robot within the environment (iCub [449],
ISAC [269], MACSi [257]). Along the same lines, Kismet and RoboCog [34] use a map built on the camera
reference frame to maintain information about recently perceived regions of interest.
30 Iuliia Kotseruba, John K. Tsotsos

Since, by definition, working memory is a relatively small temporary storage, for biological realism, its
capacity should be limited. However, there is no consensus on of how this should be done. For instance, in
GLAIR the contents of working memory are discarded when the agent switches to a new problem [484]. A
more common approach is to gradually remove items from the memory based on their recency or relevance in
the changing context. The CELTS architecture implements this principle by assigning an activation level to
percepts that is proportional to the emotional valence of a perceived situation. This activation level changes
over time and as soon as it falls below a set threshold, the percept is discarded [149]. The Novamente
Cognitive Engine has a similar mechanism, where atoms stay in memory as long as they build links to
other memory elements and increase their utility [192]. It is still unclear if under these conditions the size of
working memory can grow substantially without any additional restrictions.
To prevent unlimited growth, a hard limit can be defined for the number of items in memory, for example,
3-6 objects in ARCADIA [46], 4 chunks in CHREST [334] or up to 20 items in MDB [43]. Then as the new
information arrives, the oldest or the most irrelevant items would be deleted to avoid overflow. Items can
also be discarded if they have not been used for some time. The exact amount of time can vary from 4-9
seconds (EPIC [277]) to 5 sec (MIDAS [235], CERA-CRANIUM [23]) to tens of seconds (LIDA [170]). In
the Recommendation Architecture, a different approach is taken so that the limit of 3-4 items in working
memory emerges naturally from the structure of the memory system, and not the external parameter setting
[106].

7.3 Long-term memory

Long-term memory (LTM) preserves a large amount of information for a very long time. Typically, it is
divided into the procedural memory of implicit knowledge (e.g. motor skills and routine behaviors) and
declarative memory, which contains (explicit) knowledge. The latter is further subdivided into semantic
(factual) and episodic (autobiographical) memory.
The dichotomies between the explicit/implicit and the declarative/procedural dimensions of long-term
memories are usually merged. One of the few exceptions is CLARION, where procedural and declarative
memories are separate and both subdivided into an implicit and explicit component. This distinction is pre-
served on the level of knowledge representation: implicit knowledge is captured by distributed sub-symbolic
structures like neural networks, while explicit knowledge has a transparent symbolic representation [507].
Long-term memory is a storage for innate knowledge that enables operation of the system, therefore
almost all architectures implement procedural and/or semantic memory. The procedural memory contains
knowledge about how to get things done in the task domain. In symbolic production systems, procedural
knowledge is represented by a set of if-then rules preprogrammed or learned for a particular domain (3T
[160], 4CAPS [562], ACT-R [321], ARDIS [353], EPIC [279], SAL [225], Soar [330], APEX [173]). Other
variations include sensory-motor schemas (ADAPT [49]), task schemas (ATLANTIS [179]) and behavioral
scripts (FORR [138]). In emergent systems, procedural memory may contain sequences of state-action pairs
(BECCA [440]) or ANNs representing perceptual-motor associations (MDB [454]).
Semantic memory stores facts about the objects and relationships between them. In the architectures
that support symbolic reasoning, semantic knowledge is typically implemented as a network-like ontology,
where nodes correspond to concepts and links represent relationships between them (Disciple [52], MIDAS
[104], Soar [330], CHREST [333]). In emergent architectures factual knowledge is represented as patterns of
activity within the network (BBD [296], SHRUTI [487], HTM [319], ART [86]).
Episodic memory stores specific instances of past experience. These can later be reused if a similar
situation arises (MAX [300], OMAR [121], iCub [569], Ymir [536]). However, these experiences can also be
A Review of 40 Years in Cognitive Architecture Research Core Cognitive Abilities and Practical Applications 31

exploited for learning new semantic or procedural knowledge. For example, CLARION saves action-oriented
experiences as ”input, output, result” and uses them to bias future behavior [512]. Similarly, BECCA stores
sequences of state-action pairs to make predictions and guide the selection of system actions [440], and
MLECOG gradually builds 3D scene representation from perceived situations [258]. MAMID [239] saves
the past experience together with the specific affective connotations (positive or negative), which affect
the likelihood of selecting similar actions in the future. Other examples include R-CAST [38], Soar [395],
Novamente [327] and Theo [374].

7.4 Global memory

Despite the evidence for the distinct memory systems, some architectures do not have separate representa-
tions for different kinds of knowledge or short- vs long-term memory, and instead, use a unified structure
to store all information in the system. For example, CORTEX and RoboCog use an integrated, dynamic
multi-graph object which can represent both sensory data and high-level symbols describing the state of
the robot and the environment [442,80]. Similarly, AIS implements a global memory, which combines a
knowledge database, intermediate reasoning results and the cognitive state of the system [217]. DiPRA uses
Fuzzy Cognitive Maps to represent goals and plans [417]. NARS represents all empirical knowledge, regard-
less of whether it is declarative, episodic or procedural, as formal sentences in Narcese [582]. Similarly, in
some emergent architectures, such as SASE [590] and ART [86], the role of neurons as working or long-term
memory is dynamic and depends on whether the neuron is firing.
Overall, research on memory in the cognitive architectures mainly concerns its structure, representation
and retrieval. Relatively little attention has been paid to the challenges associated with maintaining a large-
scale memory store since both the domains and the time spans of the intelligent agents are typically limited.
In comparison, early estimates of the capacity of the human long-term memory are within 1.5 gigabits or
on the order of 100K concepts [311], while more recent findings suggest that human brain capacity may be
orders of magnitude higher [40]. However, scaling naı̈ve implementations even to the lowest estimated size
of human memory likely will not be possible despite the increase in available computational power. Thus
alternative solutions include tapping into existing methods for large-scale data management and improving
the retrieval algorithms. Both avenues have been explored, the former by Soar and ACT-R, which used
PostgreSQL relational database to load concepts and relations from WordNet [118,126], and the latter by
the Companions architecture [165]. Alternatively, SPA supports a biologically plausible model of associative
memory capable of representing over 100K concepts in WordNet using a network of spiking neurons [111].

8 Learning

Learning is the capability of a system to improve its performance over time. Ultimately, any kind of learning is
based on experience. For example, a system may be able to infer facts and behaviors from the observed events
or from results of its own actions. The type of learning and its realization depend on many factors, such as
design paradigm (e.g. biological, psychological), application scenario, data structures and the algorithms used
for implementing the architecture, etc. However, we will not attempt to analyze all these aspects given the
diversity and number of the cognitive architectures surveyed. Besides, not all of these pieces of information
can be easily found in the publications. Thus, a more general summary is preferable, where types of learning
are defined following taxonomy by Squire ([496]). Learning is divided into declarative or explicit knowledge
acquisition and non-declarative, which includes perceptual, procedural, associative and non-associative types
of learning. Figure 9 shows a visualization of these types of learning for all cognitive architectures.
32 Iuliia Kotseruba, John K. Tsotsos

e
iv

iv

iv
t

t
g ia

g ia

g ia
im s ve

im s ve

im s ve
asoceptu e

asoceptu e

asoceptu e
pr nasiati al

pr nasiati al

pr nasiati al
nosocdural

no oc ural

nosocdural
v

v
in oc

in oc

in oc
pr rcerati

pr rcerati

pr rcerati
d
pecla

pecla

pecla
s
de

de

de
Disciple iCub ART
MusiCog MACSi BBD
GLAIR Novamente MDB
AIS CoSy SPA
Companions ACT-R HTM
ERE CLARION MicroPsi
ICARUS NARS Darw. Neurodynamics
MAX CHREST Shruti
PRODIGY GMU-BICA SASE
R-CAST RCS DAC
Teton Soar BECCA
Theo CogPrime Leabra
APEX DIARC Recommendation
COGNET DSO Subsumption
EPIC CHARISMA
Homer Pogamut c) emergent
IMPRINT SAL
MAMID Xapagy
MIDAS LIDA
OMAR CARACaS
OSCAR CSE
PRS CoJACK
DiPRA
a) symbolic FORR
REM
Ymir
CELTS
ADAPT
CORTEX
RoboCog
ASMO
MLECOG
ISAC
PolyScheme
RALPH
Sigma
Kismet
ARS/SiMA
DUAL
3T
ARCADIA
ARDIS
ATLANTIS
CAPS
CERA-CRANIUM
Copycat/Metacat
MIDCA
STAR

b) hybrid
Fig. 9 A diagram summarizing types of learning implemented in the cognitive architectures. Symbolic, hybrid and emergent
architectures are grouped and presented in different columns. The architectures are arranged according to the similarity between
the implemented learning types. A version of the diagram with an alphabetical ordering of the architectures is also available on
the project website.
A Review of 40 Years in Cognitive Architecture Research Core Cognitive Abilities and Practical Applications 33

8.1 Perceptual learning

Although many systems use pre-learned components for processing perceptual data, such as object and face
detectors or classifiers, we do not consider these here. Perceptual learning applies to the architectures that
actively change the way sensory information is handled or how patterns are learned on-line. This kind of
learning is frequently performed to obtain implicit knowledge about the environment, such as spatial maps
(RCS [477], AIS [216], MicroPsi [33]), clustering visual features (HTM [292], BECCA [436], Leabra [405])
or finding associations between percepts. The latter can be used within the same sensory modality as in the
case of the agent controlled by Novamente engine, which selects a picture of the object it wants to get from
a teacher [188]. Learning can also occur between different modalities, for instance, the robot based on the
SASE architecture learns the association between spoken command and action [591] and Darwin VII (BBD)
learns to associate taste value of the blocks with their visual properties [132].

8.2 Declarative learning

Declarative knowledge is a collection of facts about the world and various relationships defined between
them. In many production systems such as ACT-R and others, which implement chunking mechanisms
(SAL [260], CHREST [476], CLARION [513]), new declarative knowledge is learned when a new chunk is
added to declarative memory (e.g. when a goal is completed). Similar acquisition of knowledge has also been
demonstrated in systems with distributed representations. For example, in DSO knowledge can be directly
input by human experts or learned as contextual information extracted from labeled training data [388].
New symbolic knowledge can also be acquired by applying logical inference rules to already known facts
(GMU-BICA [458], Disciple [51], NARS [579]). In many biologically inspired systems learning new concepts
usually takes the form of learning the correspondence between the visual features of the object and its name
(iCub [123], Leabra [405], MACsi [257], Novamente [188], CoSy [210], DIARC [612]).

8.3 Procedural learning

Procedural learning refers to learning skills, which happens gradually through repetition until the skill
becomes automatic. The simplest way of doing so is by accumulating examples of successfully solved problems
to be reused later (e.g. AIS [216], R-CAST [153], RoboCog [343]). For instance, in a navigation task, a
traversed path could be saved and used again to go between the same locations later (AIS [216]). Obviously,
this type of learning is very limited and further processing of accumulated experience is needed to improve
efficiency and flexibility.
Explanation-based learning (EBL) [370] is a common technique for learning from experience found in
many architectures with symbolic representation for procedural knowledge (PRODIGY [144], Teton [558],
Theo [558], Disciple [51], MAX [300], Soar [309], Companions [175], ADAPT [47], ERE [272], REM [381],
CELTS [148], RCS [4]). In short, it allows generalization of the explanation of a single observed instance into
a general rule. However, EBL does not extend the problem domain, but rather makes solving problems more
efficient in similar situations. A known drawback of this technique is that it can result in too many rules
(also known as the ”utility problem”), which may slow down the inference. To avoid the explosion in the
model knowledge, various heuristics can be applied, e.g. adding constraints on events that a rule can contain
(CELTS [148]) or eliminating low-use chunks (Soar [275]). Although EBL is not biologically inspired, it has
been shown that in some cases human learning may exhibit EBL-like behavior [561].
34 Iuliia Kotseruba, John K. Tsotsos

Symbolic procedural knowledge can also be obtained by inductive inference (Theo [374], NARS [491]),
learning by analogy (Disciple [527], NARS [491]), behavior debugging (MAX [300]), probabilistic reasoning
(CARACaS [249]), correlation and abduction (RCS [4]) and explicit rule extraction (CLARION [516]).

8.4 Associative learning

Associative learning is a broad term for decision-making processes influenced by reward and punishment.
In behavioral psychology, it is studied within two major paradigms: classical (Pavlovian) and instrumental
(operant) conditioning. Reinforcement learning (RL) and its variants, such as temporal difference learning,
Q-learning, Hebbian learning, etc., are commonly used in computational models of associative learning.
Furthermore, there is substantial evidence that error-based learning is fundamental for decision-making and
motor skill acquisition [233,390,479].
The simplicity and efficiency of reinforcement learning make it one of the most common techniques with
nearly half of all the cognitive architectures using it to implement associative learning. The advantage of
RL is that it does not depend on the representation and can be used in the symbolic, emergent or hybrid
architectures. One of the main uses of this technique is developing adaptive behavior. In systems with
symbolic components it can be accomplished by changing the importance of actions and beliefs based on
their success/failure (e.g. RCS [9], NARS [491], REM [381], RALPH [399], CHREST [476], FORR [198],
ACT-R [84], CoJACK [433]).
In the hybrid and emergent systems, reinforcement learning can establish associations between states
and actions. One application is sensorimotor reconstruction. The associations are often established in two
stages: a ”motor babbling” stage, during which the system performs random actions to accumulate data,
followed by the learning stage where a model is built using the accrued experiences (ISAC [269], BECCA
[440], iCub [365], DiPRA [415], CSE [222]). Associative learning can be applied when analytical solutions are
hard to obtain as in the case of soft-arm control (ISAC [269]) or when smooth and natural looking behavior
is desired (e.g. Segway platform control using BBD [295]).

8.5 Non-associative learning

Non-associative learning, as the name suggests, does not require associations to link stimuli and responses
together. Habituation and sensitization are commonly identified as two types of non-associative learning.
Habituation describes gradual reduction in the strength of response to repeated stimuli. The opposite process
occurs during sensitization, i.e. repeated exposure to stimuli causes increased response. Because of their
simplicity these types of learning are considered a prerequisite for other forms of learning. For instance,
habituation filters out irrelevant stimuli and helps to focus on important stimuli [428], especially in the
situations when positive or negative rewards are absent [587].
Most of the work to date in this area has been dedicated to the habituation in the context of social
robotics and humancomputer interaction to achieve adaptive and realistic behavior. For instance, the ASMO
architecture enables the robot to ignore irrelevant (but salient) stimuli. During the tracking task, the robot
may be easily distracted by fast motion in the background. Habituation learning (implemented as a boost
value attached to motion module) allows it to focus on the stimuli relevant for the task [393]. In Kismet
[67] and iCub [449] habituation (realized as a dedicated habituation map) causes the robot to look at non-
preferred or novel stimuli. In LIDA and Pogamut habituation is employed to reduce the emotional response
to repeated stimuli in human-robot interaction [168] and virtual agent [182] scenarios respectively. MusiCog
A Review of 40 Years in Cognitive Architecture Research Core Cognitive Abilities and Practical Applications 35

includes habituation effects to introduce novelty in music generation by aggressively suppressing the salience
of elements in working memory that have been stored for a long time [359].
In comparison, little attention has been paid to sensitization. In ASMO, in addition to habituation,
sensitization learning allows the social robot to focus on the motion near the tracked object even though it
may be slow and not salient [393]. The SASE architecture also describes both mechanisms of non-associative
learning [236].

8.6 Priming

Priming occurs when prior exposure to stimulus affects its subsequent identification and classification. Nu-
merous experiments demonstrated the existence of priming effects with respect to various stimuli, perceptual,
semantic, auditory, etc., as well as behavioral priming [228]. In Section 4.1 we mentioned some examples of
priming in vision - spatial (STAR [293]) and feature priming (Kismet [60], ART [202]), which allow for more
efficient processing of stimuli by biasing the visual system.
Priming in cognitive architectures has been investigated both within applications and theoretically. For
example, in HTM priming is used in the context of analyzing streams of data such as spoken language.
Priming is essentially a prediction of what is more likely to happen next and may resolve ambiguities
based on those expectations [211]. Experiments with CoSy show that priming speech recognition results
in a statistically significant improvement compared to a baseline system [331]. Other instances of priming
were shown in improving perceptual categorization (ARS/SiMA [467]), skill transfer (SASE [618]), problem
solving (DUAL [289]) and question answering (Shruti [486]).
Several models of priming validated by human data were developed in the lexical domain and problem
solving. For instance, the CLARION model of positive and negative priming in lexical decision tasks models
the fact that human participants are faster at identifying sequences of related concepts (e.g. the word
”butter” preceded by the word ”bread”) [220]. Leabra captures effects of priming for the production of the
English past-tense inflection [404]. Finally, Darwinian Neurodynamics confirms priming in problem solving
by showing that people who were primed for a more efficient solution perform better in solving puzzles [156].
Priming is well studied in psychology and neuroscience [592] and two major theoretical frameworks for
modeling priming are found in cognitive architectures: spreading activation (ACT-R [532], Recommendation
Architecture [106], Shruti [486], CELTS [147], LIDA [169], ARS/SiMA [467], DUAL [414], NARS [577])
and attractor networks (CLARION [220], Darwinian Neurodynamics [156].). The current consensus in the
literature is that spreading activation has greater explanation power, but attractor networks are considered
more biologically plausible [325]. In this case, it seems that the choice of the particular paradigm may also
be influenced by the representation. For instance, spreading activation is found in the localist architectures,
where units correspond to concepts. When a concept is invoked, a corresponding unit is activated, the
activation is spread to adjacent related units, which facilitates their further use. Alternatively, in attractor
networks, a concept is represented by a pattern involving multiple units. Depending on the correlation
(relatedness) between the patterns, activating one leads to increase in activation of others. An alternative
explanation of priming is offered by SPA based on parallel constraint satisfaction [478].
We also identified 19 architectures (mostly symbolic and hybrid) that do not implement any learning. In
some areas of research, learning is not even necessary, for example, in human performance modeling, where
accurate replication of human performance data is required instead (e.g. APEX, EPIC, IMPRINT, MAMID,
MIDAS, etc.). Some of the newer architectures are still in the early development stage and may add learning
in the future (e.g. ARCADIA, SiMA, Sigma).
36 Iuliia Kotseruba, John K. Tsotsos

9 Reasoning

Reasoning, originally a major research topic in philosophy and epistemology, in the past decades has be-
come one of the focal points in psychology and cognitive sciences as well. As an ability to logically and
systematically process knowledge, reasoning can affect or structure virtually any form of human activity. As
a result, aside from the classic triad of logical inference (deduction, induction and abduction), other kinds
of reasoning are now being considered, such as heuristic, defeasible, analogical, narrative, moral, etc.
Predictably, all cognitive architectures are concerned with practical reasoning, whose end goal is to
find the next best action and perform it, as opposed to the theoretical reasoning that aims at establishing
or evaluating beliefs. There is also a third option, exemplified by the Subsumption architecture, which
views reasoning about actions as an unnecessary step [75] and instead pursues physically grounded action
[77]. Granted, one could still argue that a significant amount of reasoning and planning is required from a
designer to construct a grounded system with non-trivial capabilities. Otherwise, in the context of cognitive
architectures, reasoning is primarily mentioned with regards to planning, decision-making and learning, as
well as perception, language understanding and problem-solving.
One of the main challenges a human, and consequently any human-level intelligence, faces regularly is
acting based on insufficient knowledge or ”making rational decisions against a background of pervasive ig-
norance” [422]. In addition to sparse domain knowledge, there are limitations in terms of available internal
resources, such as information processing capacity. Furthermore, external constraints, such as real-time exe-
cution, may also be introduced by the task. It should be noted that even in the absence of these restrictions,
reasoning is by no means trivial. Consider, for instance, the complex machinery used by Copycat and Metacat
to model analogical reasoning in a micro-domain of character strings [232].
So far, the most consolidated effort has been spent on overcoming the insufficient knowledge and/or re-
sources problem in continuously changing environments. In fact, this is the goal of general-purpose reasoning
systems such as the Procedural Reasoning System (PRS), the Non-Axiomatic Reasoning System (NARS),
OSCAR and the Rational Agent with Limited-Performance Hardware (RALPH). PRS is one of the earliest
instances of the BDI (belief-desire-intention) model. In each reasoning cycle, it selects a plan matching the
current beliefs, adds it to the intention stack and executes it. If new goals are generated during the execu-
tion, new intentions are created [184]. NARS approaches the issue of insufficient knowledge and resources
by iteratively reevaluating existing evidence and adjusting its solutions accordingly. This is possible with
Non-Axiomatic Logic, which associates truth values with each statement thus allowing agents to express
their confidence in the belief, a feature not available in PRS [207]. RALPH tackles complex goal-driven be-
havior in complex domain using decision-theoretic approach. Thus, computation is synonymous with action
and action with the highest utility value (based on its expected effect) is always selected [451]. The OSCAR
architecture explores defeasible reasoning, i.e. reasoning which is rationally compelling but not deductively
valid [291]. This is a more accurate representation of everyday reasoning, where knowledge is sparse, tasks
are complex and there are no certain criteria for measuring success.
All the architectures mentioned above aim at creating rational agents with human-level intelligence but
they do not necessarily try to model human reasoning processes. This is the goal of architectures such as ACT-
R, Soar, DUAL and CLARION. In particular, Human Reasoning Module implemented in ACT-R works under
the assumption that human reasoning is probabilistic and inductive (although deductive reasoning may still
occur in the context of some tasks). This is demonstrated by combining a deterministic rule-based inference
mechanism with long-term declarative memory with properties similar to human memory, i.e. incomplete
and inconsistent knowledge. Together, the uncertainty inherent in the knowledge and production rules enable
human-like reasoning behavior, which has been confirmed experimentally [396]. Similarly, symbolic and sub-
A Review of 40 Years in Cognitive Architecture Research Core Cognitive Abilities and Practical Applications 37

symbolic mechanisms in CLARION [220] and DUAL [290] architectures are used to model and explain
various psychological phenomena associated with deductive, inductive, analogical and heuristic reasoning.
As discussed in Section 3, the question of whether higher human cognition is inherently symbolic or
not is still unresolved. The symbolic and hybrid architectures considered so far, view reasoning mainly as a
symbolic manipulation. Given that in emergent architectures the information is represented by the weights
between individual units, what kind of reasoning, if any, can they support? It appears, based on the cognitive
architecture literature, that many of the emergent architectures, e.g. ART, HTM, DAC, BBD, BECCA,
simply do not address reasoning, although they certainly exhibit complex intelligent behavior. On the other
hand, since it is possible to establish a correspondence between the neural networks and logical reasoning
systems [355], then it should also be possible to simulate symbolic reasoning via neural mechanisms. Indeed,
several architectures demonstrate symbolic reasoning and planning (SASE, SHRUTI, MicroPsi). To date,
one of the most successful implementations of symbolic reasoning (and other low- and high-level cognitive
phenomena) in a neural architecture is represented by SPA [430]. Architectures like this also raise interesting
questions whether it makes sense to try and segregate symbolic parts of cognition from sub-symbolic. As
most existing cognitive architectures represent a continuum from purely symbolic to connectionist, the same
may be true for the human cognition as well.

10 Metacognition

Metacognition [161], intuitively defined as ”thinking about thinking”, is a set of abilities that introspec-
tively monitor internal processes and reason about them15 . There has been a growing interest in developing
metacognition for artificial agents, both due to its essential role in human experience and practical neces-
sity for identifying, explaining and correcting erroneous decisions. Approximately pne-third of the surveyed
architectures, mainly symbolic or hybrid ones with a significant symbolic component, support metacogni-
tion with respect to decision-making and learning. Here we will focus on three most common metacognitive
mechanisms, namely self-observation, self-analysis and self-regulation.
Self-observation allows gathering data pertaining to the internal operation and status of the system.
The data commonly includes the availability/requirements of internal resources (AIS [212], COGNET [613],
Soar [306]) and current task knowledge with associated confidence values (CLARION [508], CoJACK [146],
Companions [176], Soar [306]). In addition, some architectures support a temporal representation (trace) of
the current and/or past solutions (Companions [176], Metacat [259], MIDCA [115]).
Overall, the amount and granularity of the collected data depend on the further analysis and application.
In decision-making applications, after establishing the amount of available resources, conflicting requests for
resources, discrepancies in the conditions, etc., the system may change the priorities of different tasks/plans
(AIS [212], CLARION [508], COGNET [613], CoJACK [146], MAMID [244], PRODIGY [565], RALPH [399]).
The trace of the execution and/or recorded solutions for the past problems can improve learning. For example,
repeating patterns found in past solutions may be exploited to reduce computation for similar problems in
the future (Metacat [349], FORR [143], GLAIR [484], Soar [306]). Traces are also useful for detecting and
breaking out of repetitive behavior (Metacat [349]), as well as stopping learning if the accuracy does not
improve (FORR [143]).
Although in theory many architectures support metacognition, its usefullness has only been demonstrated
in limited domains. For instance, Metacat applies metacognition to analogical reasoning in a micro-domain
15 Due to scope limitations we did not examine other concepts related to metacognition. Discussions of the relationship between

self-awareness and metacognition, the notion of ’self’, consciousness and the appropriateness of using these psychological terms
in computational models can be found in [107, 108].
38 Iuliia Kotseruba, John K. Tsotsos

of strings of characters (e.g. if abc → abd; mrrjjj → ?). Self-watching allows the system to remember past
solutions, compare different answers and justify its decisions [349]. In PRS, metacognition enables efficient
problem solving in playing games such as Othello [453], better control of the simulated vehicle [399] and real-
time adaptability to continuously changing environments [184]. Internal error monitoring, i.e. the comparison
between the actual and expected perceptual inputs, enables the SAL architecture to determine the success
or failure of its actions in a simple task, such as stacking blocks, and to learn better actions in the future
[571]. The metacognitive abilities integrated into the dialog system make the CoSy architecture more robust
to variable dialogue patterns and increase its autonomy [97]. The Companions architecture demonstrates the
importance of self-reflection by replicating data from the psychological experiments on how people organize,
represent and combine incomplete domain knowledge into explanations [176]. Likewise, CLARION models
human data for the Metcalfe (1986) task and replicates the Gentner and Collins (1981) experiment, both of
which involve metacognitive monitoring [519].
Metacognition is required for social cognition, particularly for the skill known in the psychological litera-
ture as a Theory of mind (ToM). ToM refers to being able to acknowledge and understand mental states of
other people, use the judgment of their mental state to predict their behavior and inform one’s own decision
making. Very few architectures support this ability. For instance, most recently, Sigma demonstrated two
distinct mechanisms for ToM using as an example for several single-stage simultaneous-move games, such as
the well-known Prisoner’s dilemma. The first mechanism is automatic as it involves probabilistic reasoning
over the trellis graph and the second is a combinatorial search across the problem space [427].
PolyScheme applies ToM to perspective taking in a human-robot interaction scenario. The robot and
human in this scenario are together in a room with two traffic cones and multiple occluding elements. The
human gives a command to move towards a cone, without specifying which one. If only one cone is visible
to the human, the robot can model the scene from the human’s perspective and use this information to
disambiguate the command [540]. Another example is reasoning about beliefs in a false belief task. In this
scenario, two agents A and B observe a cookie placed in a jar. After B leaves, the cookie is moved to another
jar. When B comes back, A is able to reason that B still believes that cookie is still in the first jar [465].
ACT-R was used to build several models of false belief and second-order false belief tasks (answering
questions of the kind ”Where does Ayla think Murat will look for chocolate?”) which are typically used
to assess whether children have a Theory of Mind [543,39]. However, of particular interest is the recent
developmental model of ToM based on ACT-R, which was subsequently implemented on a mobile robot.
Several scenarios were set up to show how ToM ability can improve the quality of interaction between the
robot and humans. For instance, in a patrol scenario both the robot and human receive the task of patrolling
the south area of the building, but as they start, the instructions are changed to head west. When the human
starts walking towards south, the robot infers that she might have forgotten about the new task and reminds
her of the new changes[542].

11 Practical applications

The cognitive architectures reviewed in this paper are mainly used as research tools and very few are
developed outside of academia. However, it is still appropriate to talk about their practical applications,
since useful and non-trivial behavior in arbitrary domains is perceived as intelligent and often used to
validate the underlying theories. We seek answers to the following questions: what cognitive abilities have
been demonstrated by the cognitive architectures and what particular practical tasks have they been applied
to?
A Review of 40 Years in Cognitive Architecture Research Core Cognitive Abilities and Practical Applications 39

After a thorough search through the publications, we identified more than 900 projects implemented
using 84 cognitive architectures. Our findings are summarized in Figure 10. It shows the cognitive abilities
associated with applications, the total number of applications for each architecture (represented by the length
of the bars) and what application categories they belong to (the types of categories and the corresponding
number of applications in each category are shown by the color and length of the bar segments respectively).

11.1 Competency areas

In order to determine the scope of the research in cognitive architectures with respect to modeling various
human cognitive abilities, we began with the list of human competencies proposed by Adams et al. [1].
This list overlaps with cognitive abilities that have been identified for the purposes of evaluation of cognitive
architectures (mentioned in the beginning of Section 3), and adds several new areas such as social interaction,
emotion, building/creation, quantitative skills and modeling self/other.
In the following analysis, we depart from the work in [1] in several aspects. First, these competency areas
are meant for evaluating AI and it is expected that each area will have a set of associated scenarios. As a
result we are unable to evaluate these concepts directly, we instead will discuss which of these competency
areas are represented in existing practical applications based on the published papers. Second, the authors of
[1] define subareas for each competency area (e.g. tactical, strategic, physical and social planning). Although
in the following sections we will review some of these subareas, they are not shown in the diagram in
Figure 10. Perception, memory, attention, learning, reasoning, planning and motivation (corresponding to
internal factors in Section 6 on action selection) are covered in the previous sections and will not be repeated
here. Actuation (which we simply interpret as an ability to send motor commands both in real and simulated
environments) and social interaction will be discussed in the following subsections on practical applications in
robotics (Section 11.2.2), natural language processing (Section 11.2.5) and human-robot interaction (Section
11.2.4). Below we briefly examine the remaining competency area titled ”Building/creation” (which we
changed to ”Creativity”)16 .
The creativity competency area in [1] involves also physical construction. Since demonstrated skills
in physical construction are presently limited to toy problems, such as stacking blocks in a defined work
area using a robotic arm (Soar [310]) or in simulated environments (SAL [571]), below we will focus on
computational creativity.
Adams et al. [1] list forming new concepts and learning declarative knowledge (covered in Section 8)
within the ”creation” competency area. While it can be argued that any kind of learning leads to creation
of new knowledge or behavior, in this subsection we will focus more specifically on computational creativity
understood as computationally generated novel and appropriate results that would be regarded as creative
by humans[129,501]. Examples of such results are artifacts with artistic value or innovations in scientific
theories, mathematical proofs, etc. Some computational models of creativity also attempt to capture and
explain the psychological and neural processes underlying human creativity.
In the domain of cognitive architectures creativity remains a relatively rare feature, for instance, only 7
out of 84 architectures on our list support creative abilities. They represent many directions in the field of
16 Quantitative skills, i.e. the ability to use or manipulate quantitative information, are underrepresented in cognitive architec-
tures and therefore are excluded from our analysis. For completeness, we mention few implemented examples: counting moving
and stationary objects in a simulated environment (GLAIR [462]), computing a sum of two values (SPA [134]), identifying and
counting the number of surrounding faces in a room (DIARC [475]) and finger counting (iCub [123]). Psychological experiments
involving quantitative skills inlcude: counting tasks investigated by ACT-R [409] and CLARION [518], cognitive arithmetic
model for addition and subtraction (Soar [583]), models of multi-column subtraction used to investigate the hierarchical prob-
lem representation and solving (Teton [559], ICARUS [316] and Soar [445]).
40 Iuliia Kotseruba, John K. Tsotsos

n
n

le motionn ion

io
le motionn ion

io

ity ra g
ity ra g

ct
tiv te nin
ct
tiv te nin

m tua tiolect
m tua tiolect

ea l in o
m tionion n
ea l in o
m tionion n

g
g

cr cia eas
ac entptio
cr cia eas

ac iv e
ac entptio

a c iv e

so tarnin
so tarnin

re n i y
m asong
re n i y

ot s
m asong
ot s

ar r
ar r

a
a

at rce
at rce

e
e

pe
t
pe
t
ACT-R OSCAR
ART Ymir
Soar RALPH
HTM SAL
Leabra CERA-CRANIUM
CLARION Teton
RCS ARCADIA
SPA ASMO
EPIC ERE
IMPRINT CORTEX
LIDA DiPRA
Disciple CELTS
MIDAS MAX
DIARC MicroPsi
MDB CHARISMA
ICARUS CSE
Companions Copycat/Metacat
Polyscheme Darwinian Neurodynamics
OMAR MIDCA
ISAC MLECOG
PRODIGY MusiCog
DUAL Theo
R-CAST Homer
Sigma STAR
FORR Xapagy
SASE ARDIS
CAPS
DAC
Pogamut
CARACaS
BBD
CHREST
AIS
PRS
iCub
CoSy
GLAIR
Shruti robotics
Subsumption
Kismet psychological experiments
NARS games and puzzles
GMU-BICA misc
MACSi
3T categorization and clustering
CoJACK computer vision
REM
BECCA HPM
COGNET HRI/HCI
Recommendation virtual agents
ARS/SiMA
CogPrime NLP
MAMID
Novamente
RoboCog
ADAPT
DSO
APEX
ATLANTIS

Fig. 10 A graphical representation of the practical applications of the cognitive architectures and corresponding competency
areas defined in [1]. The plot is split into two columns for better readability (the right column being the continuation of the left
one). Within each column the table shows the implemented competency areas (indicated by gray colored cells). The stacked bar
plots show the practical applications: the length of the bar represents the total number of the practical applications found in
the publications and colored segments within each bar represent different categories (see legend) and their relative importance
(calculated as a proportion of the total number of applications implementing these categories). The architectures are sorted
by the total number of their practical applications. For short descriptions of the projects, citations and statistics for each
architecture refer to the interactive version of this diagram on the project website.
A Review of 40 Years in Cognitive Architecture Research Core Cognitive Abilities and Practical Applications 41

computational creativity such as behavioral improvisation, music generation, creative problem solving and
insight. An in-depth discussion of these problems is beyond the scope of this survey, but we will briefly
discuss the contributions in these areas.
AIS and MusiCog explore improvisation in story-telling and music generation respectively. AIS enables
improvisational story-telling by controlling puppets that collaborate with humans in Virtual Theater. The
core plot of the story is predefined, as well as a set of responses and emotions for each puppet. These
pre-recorded responses are combined with individual perceptions, knowledge base, state information and
personality of the puppets [214]. Since puppets react to both the human player and each other’s actions,
they demonstrate a wide range of variations on the basic plots. Such directed improvisation transforms
a small set of abstract directions (e.g. be playful, curious, friendly) into novel and engaging experiences
[215]. A different strategy is followed by MusiCog. It does not start with predefined settings, but instead
listens to music samples and attempts to capture their structure. After learning is completed, the acquired
structures can be used to generate novel musical themes as demonstrated on Bach BWC 1013 [360]. Another
example of synthetic music generation is Roboser, a system comprised of a mobile robot controlled by the
DAC architecture and a composition engine which converts the behavioral and control states of the robot to
sounds. Together, the predefined transformations and non-deterministic exploratory behavior of the robot
result in novel musical structures [344].
Other architectures attempt to replicate the cognitive processes of human creativity rather than its results.
One of the early models, Copycat (extended later to Metacat), examined psychological processes responsible
for analogy-making and paradigm shift (also known as the ”Aha!” phenomenon) in problem solving [348]. To
uncover deeper patterns and find better analogies, Copycat performs a random micro-exploration in many
directions. A global parameter controls the degree of randomness in the system and allows it to pursue
both obvious and peculiar paths. Multiple simulations in a microdomain of character strings show that such
approach can result in rare creative breakthroughs [232].
CLARION models insight in a more cognitively plausible way and is validated on human experimental
data. The sudden emergence of a potential solution is captured through interaction between the implicit
(distributed) and explicit (symbolic) representations [511]. For example, explicit memory search provides
stereotypical semantic associations, while implicit memory allows connecting remote concepts, hence has
more creative potential. During the iterative search, multiple candidates are found through implicit search
and are added into explicit knowledge. Simulations confirm that the stochasticity of implicit search improves
chances of solving the problem [219].
Likewise, Darwinian Neurodynamics (DN) models insight during problem solving. In order to arrive to a
solution, the system switches from analytical problem solving based on past experience to implicit generation
of new hypotheses via evolutionary search. At the core of the DN model, a population of attractor networks
stores the hypotheses which are selected for reproduction based on the fitness criteria. If the initial search
through known answers in memory fails, the evolutionary search generates new candidates via mutations.
The model of insight during solving a four-tree problem is compared to human behavioral data, accurately
demonstrating the effects of priming and prior experience and the size of working memory on problem solving
abilities [156].
Last but not least is the spiking neuron model of Remote Associates Test (RAT) often used in creativity
research. However, the focus is not on creativity per se, but rather on a biologically realistic representation
for the words in the RAT test and associations between them using the Semantic Pointer Architecture (SPA)
[200].
42 Iuliia Kotseruba, John K. Tsotsos

11.2 Application categories

We identified ten major categories of applications, namely human performance modeling (HPM), games and
puzzles, robotics, psychological experiments, natural language processing (NLP), human-robot and human-
computer interaction (HRI/HCI), computer vision, categorization and clustering, virtual agents and miscel-
laneous, which included projects not related to any major group but not numerous enough to be separated
into a group of their own. Such grouping of projects emphasizes the application aspect of each project,
although the goal of the researchers may have been different.
Note that some applications are associated with more than one category. For example, Soar has been
used to play board games with a robotic arm [286], which is relevant to both robotics and games and puzzles.
Similarly, the ACT-R model of Tower of Hanoi evaluated against the human fMRI data [14] belongs to games
and puzzles and to psychological experiments.

11.2.1 Psychological Experiments

The psychological experiments category is the largest, comprising of more than one third of all applications.
These include replications of numerous psychophysiological, fMRI and EEG experiments in order to demon-
strate that cognitive architectures can adequately model human data or give reasonable explanations for
existing psychological phenomena. If the data produced by the simulation matches the human data in some
or most aspects, it is taken as an indication that a given cognitive architecture can to some extent imitate
human cognitive processes.
Most of the experiments in our list investigate psychological phenomena related to memory, perception,
attention and decision making. However, there is very little repetition among the specific studies that were
replicated by the cognitive architectures. Notably, Tower of Hanoi (and its derivative, Tower of London) is
the only task reproduced by a handful of architectures (although not on the same human data). The task
itself requires movement of a stack of disks from one rod to another and is frequently used in psychology
to study problem solving and skill learning. Similarly, in various cognitive architectures ToH/ToL tasks are
used to evaluate strategy acquisition (Soar [450], Teton [558], CAPS [266], CLARION [517], ICARUS [318],
ACT-R [16]).
Given the importance of memory for even the most basic tasks, many phenomena related to different
types of memory were also examined. To name a few, the following experiments were conducted: test of the
effect of working memory capacity on the syntax parsing (CAPS [265]), N-back task testing spatial working
memory (ACT-R [294]), reproduction of the Morris water maze test on a mobile robot to study the formation
of episodic and spatial memory (BBD [297]), effect of priming on the speed of memory retrieval (CHREST
[313]), effect of anxiety on recall (DUAL [157]), etc.
Attention has been explored both relative to perception and for general resource allocation. For example,
models have been built to explain the well-known phenomena of inattentional blindness (ARCADIA [73]) and
attentional blink (LIDA [338]). In addition, various dual task experiments were replicated to study and model
the sources of attention limitation which reduces human ability to perform multiple tasks simultaneously.
For example, CAPS proposes a functional account of attention allocation based on a dual-task experiment
involving simultaneous sentence comprehension and mental rotation citeJust2001, ACT-R models effects of
sleep loss on sustained attention performance, which requires tracking a known location on the monitor and
a reaction task [206]. Similar experiments have been repeated using EPIC [281,278].
Multiple experiments relating perception, attention, memory and learning have been simulated using
the CHREST architecture in the context of playing chess. These include investigations of gaze patterns of
novice and expert players [312], effects of ageing on chess playing related to the reduced capacity of working
A Review of 40 Years in Cognitive Architecture Research Core Cognitive Abilities and Practical Applications 43

memory and decreased perceptual abilities [494] and effects of expertise, presentation time and working
memory capacity on the ability to memorize random chess positions [187].

11.2.2 Robotics

Nearly a quarter of all applications of cognitive architectures are related to robotics. Much effort has been
spent on navigation and obstacle avoidance, which are useful on their own and are necessary for more complex
behaviors. In particular, navigation in unstructured environments was implemented on an autonomous vehicle
(RCS [103], ATLANTIS [179]), mobile robot (Subsumption [76], CoSy [407]) and unmanned marine vehicle
(CARACaS [252]).
The fetch and carry tasks used to be very popular in the early days of robotics research as an effective
demonstration of robot abilities. Some well-known examples include a trash collecting mobile robot (3T [159])
and a soda can collecting robot (Subsumption [78]). Through a combination of simple vision techniques, such
as edge detection and template matching, and sensors for navigation, these robots were able to find the objects
of interest in unknown environments.
More recent cognitive architectures solve search and object manipulation tasks separately. Typically,
experiments involving visual search are done in very controlled environments and preference is given to
objects with bright colors or recognizable shapes to minimize visual processing, for example, a red ball
(SASE [591]) or a soda can (ISAC [268]). Sometimes markers, such as printed barcodes attached to the
object, are used to simplify recognition (Soar [368]). It is important to note that visual search in these cases
is usually a part of a more involved task, such as learning by instruction. When visual search and localization
are the end goal, the environments are more realistic (e.g. the robot controlled by CoSy finds a book on a
cluttered shelf using a combination of sensors and SIFT features [335]).
Object manipulation involves arm control to reach and grasp an object. While reaching is a relatively
easy problem and many architectures implement some form of arm control, gripping is more challenging
even in a simulated environment. The complexity of grasping depends on many factors including the type
of gripper and the properties of the object. One workaround is to experiment with grasping on soft objects,
such as plush toys (ISAC [270]). More recent work involves objects with different grasping types (objects
with handles located on the top or on a side) demonstrated on a robot controlled by DIARC [599]. Another
example is iCub adapting its grasp to cans of different sizes, boxes and a ruler [464].
A few architectures implement multiple skills for complex scenarios such as a robotic salesman (CORTEX
[442], RoboCog [441]), tutoring (DAC [572]), medical assessment (LIDA [339], RoboCog [36]), etc. Industrial
applications are represented by a single architecture - RCS, which has been used for teleoperated robotic
crane operation [336], bridge construction [59], autonomous cleaning and deburring workstation [383], and
the automated stamp distribution center for the US Postal Service [8].
The biologically motivated architectures focus on the developmental aspect of physical skills and sen-
sorimotor reconstruction. For example, a childlike iCub robot platform explores acquisition of skills for
locomotion, grasping and manipulation [5], robots Dav and SAIL learn vision-guided navigation (SASE
[591]) and ISAC learns grasping affordances [555].

11.2.3 Human Performance Modeling (HPM)

Human performance modeling is an area of research concerned with building quantitative models of human
performance in a specific task environment. The need for such models comes from engineering domains where
the space of design possibilities is too large so that empirical assessment is infeasible or too costly.
44 Iuliia Kotseruba, John K. Tsotsos

This type of modeling has been used extensively for military applications, for example, workload analysis
of Apache helicopter crew [11], modeling the impact of communication tasks on the battlefield awareness
[372], decision making in the AAW domain [615], etc. Common civil applications include models of air traffic
control task citeSeamster1993), aircraft taxi errors [595], 911 dispatch operator [209], etc.
Overall, HPM is dominated by a handful of specialized architectures, including OMAR, APEX, COGNET,
MIDAS and IMPRINT. In addition, Soar was used to implement a pilot model for large-scale distributed
military simulations (TacAir-Soar [302,261]).

11.2.4 Human-Robot and Human/Computer Interaction (HRI/HCI)

HRI is a multidisciplinary field studying various aspects of communication between people and robots. Many
of these interactions are being studied in the context of social, assistive or developmental robotics. Depending
on the level of autonomy demonstrated by the robot, interactions extend from direct control (teleoperation)
to full autonomy of the robot enabling peer-to-peer collaboration. Although none of the systems presented
in this survey are yet capable of full autonomy, they allow for some level of supervisory control ranging from
single vowels signifying direction of movement for a robot (SASE [589]) to natural language instruction (Soar
[309], HOMER [566], iCub [539]). It is usually assumed that a command is of particular form and uses a
limited vocabulary.
Some architectures also target non-verbal aspects of HRI, for example, natural turn-taking in a dialogue
(Ymir [537], Kismet [62]), changing facial expression (Kismet [64]) or turning towards the caregiver (MACsi
[19]).
An important practical application which also involves HCI is in building decision support systems, i.e.
intelligent assistants which can learn from and cooperate with experts to solve problems in complex do-
mains. One such domain is intelligence analysis, which requires mining large amounts of textual information,
proposing a hypothesis in search of evidence and reevaluating hypotheses in view of new evidence (Disciple
[526]). Other examples include time- and resource-constrained domains such as emergency response planning
(Disciple [525], NARS [491]), medical diagnosis (OSCAR [423]), military operations in urban areas (R-CAST
[154]) and air traffic control (OMAR [120]).

11.2.5 Natural Language Processing (NLP)

Natural language processing (NLP) is a broad multi-disciplinary area which studies understanding of written
or spoken language. In the context of cognitive architectures, many aspects of NLP have been considered,
from low-level auditory perception, syntactic parsing and semantics to conversation in limited domains.
As has been noted in the section on perception, there are only a few models of low-level auditory per-
ception. For instance, models based on Adaptive Resonance Theory (ART), namely ARTPHONE [202],
ARTSTREAM [201], ARTWORD [205] and others, have been used to model perceptual processes involved
in speech categorization, auditory streaming, source segregation (also known as the ”cocktail party problem”)
and phonemic integration.
Otherwise, most research in NLP is related to investigating aspects of the syntactic and semantic pro-
cessing of textual data. Some examples include anaphora resolution (Polyscheme [301], NARS [283], DIARC
[597]), learning English passive voice (NARS [283]), models of syntactic and semantic parsing (SPA [498],
CAPS [265]) and word sense disambiguation (SemSoar and WordNet [263]).
One of the early models for real-time natural language processing, NL-Soar, combined syntactic knowledge
and semantics of simple instructions for the immediate reasoning and tasks in blocks world [326]. A more
recent model (currently used in Soar) is capable of understanding commands, questions, syntax and semantics
A Review of 40 Years in Cognitive Architecture Research Core Cognitive Abilities and Practical Applications 45

in English and Spanish [330]. A complete model of human reading which simulated gaze patterns, sequential
and chronometric characteristics of human readers, semantic and syntactic analysis of sentences, recall and
forgetting was built using the CAPS architecture [531].
More commonly, off-the-shelf software is used for speech recognition and parsing, which helps to achieve
a high degree of complexity and realism. For instance, a salesman robot (CORTEX [81]) can understand and
answer questions about itself using a Microsoft Kinect Speech SDK. The Playmate system based on CoSy
uses dedicated software for speech processing [331] and can have a meaningful conversation in a subset of
English about colors and shapes of objects on the table. The FORR architecture uses speech recognition for
the task of ordering books from the public library by phone. The architecture, in this case, increases the
robustness of the automated speech recognition system based on an Olympus/RavenClaw pipeline (CMU)
[142].
In general, most existing NLP systems are limited both in their domain of application and in terms of
the syntactic structures they can understand. Recent architectures, such as DIARC, aim at supporting more
naturally sounding requests like Can you bring me something to cut a tomato?, however, they are still in
the early stages of development [463].

11.2.6 Categorization and clustering

Categorization, classification, pattern recognition and clustering are common ways of extracting general
information from large datasets. In the context of cognitive architectures, these methods are useful for
processing noisy sensory data. Applications in this group are almost entirely implemented by the emergent
architectures, such as ART and HTM, which are used as sophisticated neural networks. The ART networks, in
particular, have been applied to classification problems in a wide range of domains: movie recommendations
(Netflix dataset [87]), medical diagnosis (Pima-Indian diabetes dataset [271]), fault diagnostics (pneumatic
system analysis [116]), vowel recognition (Peterson and Barney dataset [12]), odor identification [125], etc.
The HTM architecture is geared more towards the analysis of time series data, such as predicting IT failures17 ,
monitoring stocks18 , predicting taxi passenger demand [113] and recognition of cell phone usage type (email,
call, etc.) based on the pressed key pattern [363].
A few other examples from the non-emergent architectures include gesture recognition from tracking
suits (Ymir [93]), diagnosis of the failures in a telecommunications network (PRS [429]) and document
categorization based on the information about authors and citations (CogPrime [208]).

11.2.7 Computer vision

The emergent cognitive architectures are also widely applied to solving typical computer vision problems.
However, these are mainly standalone examples, such as hand-written character recognition (HTM [538,503]),
image classification benchmarks (HTM [619,342]), view-invariant letter recognition (ART [155]), texture
classification benchmarks (ART [576]), invariant object recognition (Leabra [403]), etc.
The computer vision applications that are part of more involved tasks, such as navigation in robotics,
are discussed in the relevant sections.

17 grokstream.com
18 numenta.com/htm-for-stocks
46 Iuliia Kotseruba, John K. Tsotsos

11.2.8 Games and puzzles

Playing games has been an active area of research in cognitive architectures for decades. Some of the earliest
models for playing tic-tac-toe and Eight Puzzle as a demo of reasoning and learning abilities were created
in the 1980s (Soar [305]). The goal is typically not to master the game but rather to use it as a step towards
solving similar but more complex problems. For example, Liar’s Dice, a multi-player game of chance, is used to
assess the feasibility of reinforcement learning in large domains (Soar [117]). Similarly, playing Backgammon
was used to model cognitively plausible learning in (ACT-R [460]) and tic-tac-toe to demonstrate ability to
learn from instruction (Companions [231]). Multiple two-player board games with conceptual overlap like
tic-tac-toe, the Eight Puzzle and the Five Puzzle can also be used as an effective demonstration of knowledge
transfer (e.g. Soar [286], FORR [140]).
Compared to the classic board games used in AI research since its inception, video games provide a
much more varied and challenging domain. Like board games, most video games require certain cognitive
skills from a player, thus allowing the researchers to break down the problem of solving general intelligence
into smaller chunks and work on them separately. However, the graphics, complexity and response times of
the recent video games are getting better and better, hence many games can already be used as sensible
approximations of the real world environments. In addition to that, the simulated intelligent entity is much
cheaper to develop and less prone to damage than the one embodied in a physical robot. A combination of
these factors makes video games a very valuable test platform for models of human cognition.
The only drawback of using video games for research is that embedding a cognitive architecture within
it requires software engineering work. Naturally, games with open source engines and readily available mid-
dleware are preferred. One such example is the Unreal Tournament 2004 (UT2004) game, for which the
Pogamut architecture [267] serves as a middleware, making it easier to create intelligent virtual characters.
Although Pogamut itself implements many cognitive functions, it is also used with modifications by other
groups to implement artificial entities for UT2004 [379,493,112,575,557]. Other video games used in cogni-
tive architectures research are Freeciv (REM [554]), Atari Frogger II (Soar [605]), Infinite Mario (Soar [489]),
browser games (STAR [293]) and custom made games (Soar [346]). It should be noted that aside from the
playing efficiency and achieved scores the intelligent agents are also evaluated based on their believability
(e.g. 2K BotPrize Contest19 ).

11.2.9 Virtual agents

Although related to human performance modeling, the category of virtual agents is broader. While HPM
requires models which can closely model human behavior in precisely defined conditions, the goals of creating
virtual agents are more varied. Simulations and virtual reality are frequently used as an alternative to
the physical embodiment. For instance, in the military domain, simulations model behavior of soldiers in
dangerous situations without risking their lives. Some examples include modeling agents in a suicide bomber
scenario (CoJACK [145]), peacekeeping mission training (MAMID [239]), command and control in complex
and urban terrain (R-CAST [152]) and tank battle simulation (CoJACK [434]).
Simulations are also common for modeling behaviors of intelligent agents in civil applications. One of
the advantages of virtual environments is that they can provide information about the state of the agent at
any point in time. This is useful for studying the effect of emotions on actions, for example, in the social
interaction context (ARS/SiMA [466]), or in learning scenarios, such as playing fetch with a virtual dog
(Novamente [221]).
19 http://botprize.org
A Review of 40 Years in Cognitive Architecture Research Core Cognitive Abilities and Practical Applications 47

Intelligent characters with engaging personalities can also enhance user experience and increase their level
of engagement in video games and virtual reality applications. One of the examples is virtual reality drama
”Human Trials” which lets human actors participate in an immersive performance together with multiple
synthetic characters controlled by the GLAIR architecture [483,18]). A similar project called Virtual Theater
allowed users to interact with virtual characters on the screen to improvise new stories in real time. The
story unfolds as users periodically select among directional options appearing on the screen (AIS [215]).

Discussion

The main contribution of this survey is in gathering and summarizing information on a large number of
cognitive architectures influenced by various disciplines (computer science, cognitive psychology, philosophy
and neuroscience). In particular, we discuss common approaches to modeling important elements of human
cognition, such as perception, attention, action selection, learning, memory and reasoning. In our analysis,
we also point out what approaches are more successful in modeling human cognitive processes or exhibiting
useful behavior. In order to evaluate practical aspects of cognitive architectures we categorize their existing
practical applications into several broad categories. Furthermore, we map these practical applications to
a number of competency areas required for human-level intelligence to assess the current progress in that
direction. This map may also serve as an approximate indication of what major areas of human cognition
have received more attention than others. Thus, there are two main outcomes of this review. First, we
present a broad and inclusive snapshot of the progress made in cognitive architectures research over the past
four decades. Second, by documenting the wide variety of tested mechanisms that may help in developing
explanations and models for observed human behavior, we inform future research in cognitive science and in
the component disciplines that feed it. In the remainder of this section we will summarize the main themes
discussed in the present survey and pertaining issues which can guide directions for future research.
Approaches to modeling human cognition. Historically, psychology and computer science were
inspirations for the first cognitive architectures, i.e. the theoretical models of human cognitive processes and
corresponding software artifacts that allowed demonstrating and evaluating the underlying theory of mind.
Agent architectures, on the other hand, focus on reproducing a desired behavior and particular application
demands without the constraints of cognitive plausibility.
While often described as opposing, in practice the boundary between the two paradigms is not well defined
(which causes inconsistencies in the survey literature as discussed in Section 2). Despite the differences in
the theory and terminology both cognitive and agent architectures have functional modules corresponding to
human cognitive abilities and tackle the issues of action selection, adaptive behavior, efficient data processing
and storage. For example, methods of action selection widely used in robotics and classic AI, ranging from
priority queue to reinforcement learning [419], are also found in many cognitive architectures. Likewise, some
agent architectures incorporate cognitively plausible structures and mechanisms.
Biologically and neurally plausible models of human mind aim to explain how known cognitive and
behavioral phenomena arise from low-level brain processes. It has been argued that their support for inference
and general reasoning is inadequate for modeling all aspects of human cognition, although some of the latest
neuronal models demonstrate that it is far from being proven.
The newest paradigm in AI is represented by machine learning methods, particularly data-driven deep
learning, which have found enormous practical success in limited domains and even claim some biological
plausibility. The bulk of work in this area is on perception although there have been attempts at implementing
more general inference and memory mechanisms. The results of these efforts are stand-alone models that
48 Iuliia Kotseruba, John K. Tsotsos

have yet to be incorporated within a unified framework. At present these new techniques are not widely
incorporated into existing cognitive architectures.
Range of supported cognitive abilities and remaining gaps. The list of cognitive abilities and
phenomena covered in this survey is by no means exhaustive. However, none of the systems we reviewed
is close to supporting in theory or demonstrating in practice even this restricted subset, let alone a set of
identified cognitive abilities (a comprehensive survey by Carroll [92] lists nearly 3000 of them).
Most effort thus far has been dedicated to studying high-level abilities such as action selection (reactive,
deliberative and, to some extent, emotional), memory (both short and long-term), learning (declarative and
non-declarative) and reasoning (logical inference, probabilistic and meta-reasoning). Decades of work in these
areas, both in traditional AI research and cognitive architectures, resulted in creation of multiple algorithms
and data structures. Furthermore, various combinations of these approaches have been theoretically justified
and practically validated in different architectures, albeit under restricted conditions.
We only briefly cover abilities that rely on the core set of perception, action-selection, memory and learn-
ing. At present, only a fraction of the architectures have the theoretical and software/hardware foundation
to support creative problem solving, communication and social interaction (particularly in the fields of social
robotics and HCI/HRI), natural language understanding and complex motor control (e.g. grasping).
With respect to the range of abilities and realism of the scenarios there remains a significant gap between
the state-of-the-art in specialized areas of AI and the research in these areas within the cognitive architectures
domain. Arguably some of the AI algorithms demonstrate achievements that are already on par with or even
exceed human abilities (e.g. in image classification [218], face recognition [522], playing simple video games
[376]), which raises the bar for cognitive architectures.
Due to the complexity of modeling the human mind, groups of researchers have been pursuing comple-
mentary lines of investigation. Owing to tradition, work on high-level cognitive architectures is still centered
mainly on reasoning and planning in simulated environments, downplaying issues of cognitive control of
perception, symbol grounding and realistic attention mechanisms. On the other hand, embodied architec-
tures must deal with the real-world constraints and thus pay more attention to low-level perception-motor
coupling, while most of the deliberative planning and reasoning effort is spent on navigation and motor
control. There are, however, trends towards overcoming these differences as classical cognitive architectures
experiment with embodiment [377,542] or are being interfaced with robotic architectures [472] to utilize their
respective strengths.
There also remains a gap between the biologically inspired architectures and neural simulations and the
rest of the architectures. Even though these models can demonstrate how low-level neural processes give
rise to some higher-level cognitive functions, they do not have the same range and efficiency in practical
applications compared to the less theoretically restricted systems. Some exceptions exist, for example Grok
- a commercial application for IT analytics based on the biologically inspired HTM architecture or Neural
Information Retrieval System implemented as a hierarchy of ART networks for storing 2D and 3D parts
designs built for the Boeing company [495].
Finally, most of the featured architectures cannot reuse the capabilities or accumulate knowledge as they
are applied to new tasks. Instead, every new task or skill is demonstrated using a separate model, specific
set of parameters or knowledge base 20 . Examples of architectures capable of switching between various
scenarios/environments without resetting or changing parameters are sparse and are explicitly declared as
such, e.g. the SPA architecture which can perform eight unrelated tasks [134].

20 Note that the proposed visualizations demonstrate aggregated statistics and do not imply that there exists a single model

that encapsulates them all.

A Review of 40 Years in Cognitive Architecture Research Core Cognitive Abilities and Practical Applications 49

Outstanding issues and future directions. One of the outcomes of this survey, besides summarizing
past and present attempts at modeling various human cognitive abilities, is highlighting the problems needing
further research via interactive visualization. We also reviewed the future directions suggested by scholars
in a large number of publications in search of problems deemed important. We present some of the main
themes below, excluding architecture- and implementation-specific issues.
Adequate experimental validation. A large number of papers end in a call for thorough experimental
testing of the given architecture in more diverse, challenging and realistic environments [159,438,394] and
real-world situations [511] using more elaborate scenarios [447] and diverse tasks [226]. This is by far the
most pressing and long-standing issue (found in papers from the early 90s [358] up until 2017 [156]) which
remains largely unresolved. The data presented in this survey (e.g. Section 11.2 on practical applications)
shows that, by and large, the architectures (including those vying for AGI) have practical achievements
only in several distinct areas and in highly controlled environments. Likewise, common validation methods
have many limitations. For instance, replicating psychological experiments, the most frequently used in our
sample of architectures, presents the following challenges: it is usually based on a small set of data points
from human participants, the amount of prior knowledge for the experiment is often unknown and modeled
in an ad hoc manner, pre-processing of stimuli is required in case the architecture does not posess adequate
perception, and, furthermore, the task definition is narrow and tests only a particular aspect of the ability
[262]. Overall, such restricted evaluation tells us little about the actual abilities of the system outside the
laboratory setting.
Realistic perception. The problem of perception is most acute in robotic architectures operating in
noisy unstructured environments. Frequently mentioned issues inlcude the lack of active vision [237], accurate
localization and tracking [606], robust performance under noise and uncertainty [441] and the utilization
of context information to improve detection and localization [97]. More recently, deep learning is being
considered as a viable option for bottom-up perception [257,54].
These concerns are supported by our data as well. Overall, in comparison to higher-level cognitive abilities,
the treatment of perception, and vision in particular, is rather superficial, both in terms of capturing the
underlying processes and practical applications. For instance, almost half of the reviewed architectures do not
implement any vision (Figure 5). The remaining projects, with few exceptions, either default to simulations
or operate in controlled environments.
Other sensory modalities, such as audition, proprioception and touch, often rely on off-the-shelf software
solutions or are trivially implemented via physical sensors/simulations. Multi-modal perception in the archi-
tectures that support it, is approached mechanistically and is tailored for a particular application. Finally,
bottom-up and top-down interactions between perception and higher-level cognition are largely overlooked
both in theory and practice. As we discuss in Section 5 many of the mechanisms are a by-product of other
design decisions rather than a result of deliberate efforts.
Human-like learning. Even though various types of learning are represented in the cognitive architec-
tures, there is still a need for developing more robust and flexible learning mechanisms [231,364], knowledge
transfer [226] and accumulation of new knowledge without affecting prior learning [264]. Learning is especially
important for developmental approach to modeling cognition, specifically motor skills and new declarative
knowledge (as discussed in Section 8). However, artificial development is restricted and far less efficient
because visual and auditory perception in robots is not yet on par with even the small children [97].
Natural communication. Verbal communication is the most common mode of interaction between the
artificial agents and humans. Many issues are yet to be resolved as current approaches do not possess a
sufficiently large knowledge base for generating dialogues [149] and generally lack robustness [231,596]. Few,
if any, architectures are capable of detecting the emotional state and intentions of the interlocutor [61] or
give personalized responses [186]. Non-verbal aspects of communication, such as performing and detecting
50 Iuliia Kotseruba, John K. Tsotsos

gestures and facial expressions, natural turn-taking, etc., are investigated by only a handful of architectures,
many of which are no longer in development. As a result, demonstrated interactions are heavily scripted and
often serve as voice control interfaces for supplying instructions instead of enabling collaboration between
human and the machine [385] (see also discussion in Section 11.2.5).
Autobiographic memory. Even though, memory, a necessary component of any computational model,
is well represented and well researched in the domain of cognitive architectures, episodic memory is compar-
atively under-examined [114,53,57,364]. The existence of this memory structure has been known for decades
[552] and its importance for learning, communication and self-reflection widely recognized, however episodic
memory remains relatively neglected in computational models of cognition (see Section 8). This is also true
for general models of human mind such as Soar and ICARUS, which included it relatively recently.
Currently the majority of the architectures save episodes as time-stamped snapshots of the state of the
system. This approach is suitable for agents with a short life-span and a limited representational detail but
more elaborate solutions will be needed for life-long learning and managing large-scale memory storage [308].
These issues are relevant to other types of long-term memory as well [381,224,44].
Computational performance. The problem of computational efficiency, for obvious reasons, is more
pressing in robotic architectures [180,600] and interactive applications [25,24]. However, it is also reported
for non-embodied architectures [508,263], particularly neural simulations [585,544]. Overall, time and space
complexity is glossed over in the literature, frequently described in qualitative terms (e.g. ’real-time’ without
specifics) or omitted from discussion altogether.
A related point is an issue of scale. It has been shown that scaling up existing algorithms to larger
amounts of data introduces additional challenges, such as the need for more efficient data processing and
storage. For example, the difficulties associated with maintaining a knowledge base equivalent to human
brain memory-storage capacity are discussed in Section 7. These problems are currently addressed by only
a handful of architectures, even though solving them is crucial for further development of theoretical and
applied AI.
In addition to concerns collectively expressed by the researches (and supported by the data in this survey)
below we examine two more observations: 1) the need for an objective evaluation of the results and measuring
the overall progress towards understanding and replicating human-level intelligence and 2) reproducibility
of the research in the cognitive architectures.
Comparative evaluation of cognitive architectures. The resolution of the commonly acknowledged
issues hinges on the development of objective and extensive evaluation procedures. We have already men-
tioned as one of the issues that individual architectures are validated using a disjointed and sparse set of
experiments, which cannot support claims about the abilities of the system or adequately assess its repre-
sentational power. However, comparisons between architectures and measuring overall progress of the field
entails an additional set of challenges.
Surveys are a first step towards mapping the landscape of approaches to modeling human intelligence,
but they can only identify the blank spots needing further exploration and are ill-suited for comparative
analysis. For instance, in this paper, visualizations show only the existence of any particular feature, be it
memory, learning method or perception modality. However, they cannot represent the extent of the ability
nor whether it is enabled by the architectural mechanisms or is an ad hoc implementation21 . Thus, any
comparisons between the architectures are discouraged and any such conclusions should be taken with
caveats.
A compounding factor for evaluation is that cognitive architectures comprise both theoretical and en-
gineering components, which leads to a combinatorial explosion of possible implementations. Consider, for
21 Refer to a relevant discussion in Section 7 on how working memory size is set in various architectures.
A Review of 40 Years in Cognitive Architecture Research Core Cognitive Abilities and Practical Applications 51

example, the theoretical and technical nuances of various instantiations of the global workspace concept,
e.g. ARCADIA [72], ASMO [392] and LIDA [172] or the number of planning algorithms and approaches to
dynamic action selection in Section 6. Furthermore, building a cognitive architecture is a major commit-
ment, requiring years of development to put even the basic components together (average age of projects
in our sample is ≈ 15 years), making any major changes harder with time. Thus, evaluation is crucial for
identifying more promising approaches and pruning suboptimal paths for exploration. To some extent, it has
already been happening organically via a combination of theoretical considerations and collective experience
gathered over the years. For instance, symbolic representation, once dominant, is being replaced by more
flexible emergent approaches (see Figure 2) or augmented with sub-symbolic elements (e.g. Soar, REM).
Clearly, more consistent and directed effort is needed to make this process faster and more efficient.
Ideally, proper evaluation should target every aspect of cognitive architectures using a combination of
theoretical analysis, software testing techniques, benchmarking, subjective evaluation and challenges. All
of these have already been sparingly applied to individual architectures. For example, a recent theoretical
examination of the knowledge representations used in major cognitive architectures reveals their limitations
and suggests ways to amend them [328,329]. Over the years, a number of benchmarks and challenges have
been developed in other areas of AI and a few have been already used to evaluate cognitive architectures, e.g.
2K BotPrize competition (Pogamut [181]), various benchmarks, including satellite image classification (ART
[87]), medical data analysis (ART [91]) and face recognition (HTM [503]), as well as robotics competitions
such as AAAI Robot Competition (DIARC [470] and RoboCup (BBD [520]).
Tests probing multiple abilities via series of tasks are suggested for evaluating cognitive architectures
and general-purpose AI systems. Some examples include, ”Cognitive Decathlon” [380] and I-Athlon [2], both
inspired by the Turing test, as well as Catell-Horn-Carroll model examining factors of intelligence [253] or a
battery of tests for assessing different cognitive dimensions proposed by Samsonovich et al. [457]. However,
most of the work in this area remains theoretical and has not resulted in concrete proposals ready to be
implemented or complete frameworks. For further discussion of ample literature concerning evaluation of AI
and cognitive architectures we refer the readers to a very recent and comprehensive review of these methods
in [227].
Reproducibility of the results. Overall, in the literature on cognitive architectures far more impor-
tance is given to the cognitive, psychological or philosophical aspects, while technical implementation details
are often incomplete or missing. For example, in Section 4.1 on visual processing we list architectures that
only briefly mention some perceptual capabilities but do not specify enough concrete facts for analysis. Like-
wise, justifications of using particular algorithms/representations along with their advantages and limitations
are not always disclosed. Generally speaking, lack of full technical detail compromises reproducibility of the
research.
In part, these issues can be alleviated by providing access to the software/source code, but only one-thirds
of the architectures do so. To a lesser extent this applies to architectures such as BBD, SASE, Kismet, Ymir,
Subsumption and a few others that are physically embodied and depend on a particular robotic platform.
However, the majority of architectures we reviewed have a substantial software component, releasing which
could be of benefit to the community and the researchers themselves. For instance, cognitive architectures
such as ART, ACT-R, Soar, HTM and Pogamut, are used by many researchers outside the main developing
group.
In conclusion, our hope is that this work will serve as an overview and a guide to the vast field of cognitive
architecture research as we strived to objectively and quantitatively assess the state-of-the-art in modeling
human cognition.
52 Iuliia Kotseruba, John K. Tsotsos

References

1. Sam Adams, Itmar Arel, Joscha Bach, Robert Coop, Rod Furlan, Ben Goertzel, J. Storrs Hall, Alexei Samsonovich,
Matthias Scheutz, Matthew Schlesinger, Stuart C. Shapiro, and John Sowa. Mapping the Landscape of Human-Level
Artificial General Intelligence. AI Magazine, 33(1):25–42, 2012.
2. Sam S. Adams, Guruduth Banavar, and Murray Campbell. I-athlon: Toward a Multidimensional Turing Test. AI Mag.,
31(1):78–84, 2016.
3. J. Albus and A. Barbera. Intelligent control and tactical behavior development: A long term NIST partnership with the
army. In 1st Joint Emergency Preparedness and Response/Robotic and Remote Systems Topical Meeting, 2006.
4. J. Albus, A. Lacaze, and A. Meystel. Theory and Experimental Analysis of Cognitive Processes in Early Learning. In
Proceedings of the IEEE International Conference on Systems, Man and Cybernetics, pages 4404–4409, 1995.
5. J. S. Albus. A Reference Model Architecture for Intelligent Systems Design. An introduction to intelligent and autonomous
control, pages 27–56, 1994.
6. J. S. Albus. 4D/RCS A Reference Model Architecture for Intelligent Unmanned Ground Vehicles. In Proceedings of the
SPIE 16th Annual International Symposium on Aerospace/Defense Sensing, Simulation and Controls, 2002.
7. James Albus, Roger Bostelman, Tsai Hong, Tommy Chang, Will Shackleford, and Michael Shneier. THE LAGR PROJECT
Integrating learning into the 4D/RCS Control Hierarchy. In International Conference in Control, Automation and
Robotics, 2006.
8. James S. Albus. The NIST Real-time Control System (RCS): an approach to intelligent systems research. Journal of
Experimental & Theoretical Artificial Intelligence, 9(2-3):157–174, 1997.
9. James S. Albus and Anthony J. Barbera. RCS: A cognitive architecture for intelligent multi-agent systems. Annual
Reviews in Control, 29(1):87–99, 2005.
10. James S. Albus, Hui-Min Huang, Elena R. Messina, Karl Murphy, Maris Juberts, Alberto Lacaze, Stephen B. Balakirsky,
Michael O. Shneier, Tsai Hong Hong, Harry a. Scott, Frederick M. Proctor, William P. Shackleford, John L. Michaloski,
Albert J. Wavering, Thomas R. Kramer, Nicholas G. Dagalakis, William G. Rippey, Keith a. Stouffer, and Steven Legowik.
4D/RCS: A Reference Model Architecture For Unmanned Vehicle Systems Version 2.0. In Proceedings of the SPIE 16th
Annual International Symposium on AerospaceDefense Sensing Simulation and Controls, 2002.
11. Laurel Allender. Modeling Human Performance: Impacting System Design, Performance, and Cost. In Proceedings of the
Military, Government and Aerospace Simulation Symposium, pages 139–144, 2000.
12. Heather Ames and Stephen Grossberg. Speaker normalization using cortical strip maps: a neural model for steady-state
vowel categorization. The Journal of the Acoustical Society of America, 124:3918–3936, 2008.
13. J. R. Anderson, L. M. Reder, and C. Lebiere. Working memory: activation limitations on retrieval. Cognitive psychology,
30:221–256, 1996.
14. John R. Anderson, Mark V. Albert, and Jon M. Fincham. Tracing problem solving in real time: fMRI analysis of the
subject-paced Tower of Hanoi. Journal of Cognitive Neuroscience, 17(8):1261–1274, 2005.
15. John R. Anderson, Daniel Bothell, Michael D. Byrne, Scott Douglass, Christian Lebiere, and Yulin Qin. An Integrated
Theory of the Mind. Psychological Review, 111(4):1036–1060, 2004.
16. John R. Anderson and Scott Douglass. Tower of Hanoi: Evidence for the cost of goal retrieval. Journal of Experimental
Psychology: Learning, Memory, and Cognition, 27(6):1331–1346, 2001.
17. John R Anderson and Christian Lebiere. The Newell Test for a theory of cognition. The Behavioral and Brain sciences,
26(5):587–601, 2003.
18. Josephine Anstey, Sarah Bay-Cheng, Dave Pape, and Stuart C. Shapiro. Human trials: An Experiment in Intermedia
Performance. ACM Computers in Entertainment, 5(3), 2007.
19. Salvatore M. Anzalone, Serena Ivaldi, Olivier Sigaud, and Mohamed Chetouani. Multimodal people engagement with
iCub. Biologically inspired cognitive architectures, pages 59–64, 2012.
20. Raúl Arrabales, Agapito Ledezma, and Araceli Sanchis. A cognitive approach to multimodal attention. Journal of Physical
Agents, 3(1):53–63, 2009.
21. Raul Arrabales, Agapito Ledezma, and Araceli Sanchis. CERA-CRANIUM: A Test Bed for Machine Consciousness
Research. In International Workshop on Machine Consciousness, 2009.
22. Raul Arrabales, Agapito Ledezma, and Araceli Sanchis. Towards conscious-like behavior in computer game characters. In
2009 IEEE Symposium on Computational Intelligence and Games, pages 217–224, 2009.
23. Raúl Arrabales, Agapito Ledezma, and Araceli Sanchis. Simulating Visual Qualia in the CERA-CRANIUM Cognitive
Architecture. In From Brains to Systems. Springer New York, 2011.
24. Raúl Arrabales Moreno and Araceli Sanchis de Miguel. A machine consciousness approach to autonomous mobile robotics.
In Proceedings of the 5th International Cognitive Robotics Workshop, 2006.
25. David Ash and Barbara Hayes-Roth. Using action-based hierarchies for real-time diagnosis. Artificial Intelligence, 88:317–
347, 1996.
A Review of 40 Years in Cognitive Architecture Research Core Cognitive Abilities and Practical Applications 53

26. Amal Asselman, Souhaib Aammou, and Az-Eddine Nasseh. Comparative Study of Cognitive Architectures. International
Research Journal of Computer Science, 2(9):8–13, 2015.
27. R. C. Atkinson and R. M. Shiffrin. Human Memory: A Proposed System and its Control Processes. Psychology of Learning
and Motivation - Advances in Research and Theory, 2(C):89–195, 1968.
28. Bernard J. Baars. Global workspace theory of consciousness: Toward a cognitive neuroscience of human experience.
Progress in Brain Research, 150:45–53, 2005.
29. Bernard J. Baars, Uma Ramamurthy, and Stan Franklin. How deliberate, spontaneous and unwanted memories emerge
in a computational model of consciousness. Involuntary memory, 2007.
30. Joscha Bach. Principles of Synthetic Intelligence. PhD Thesis, 2007.
31. Joscha Bach. A motivational system for cognitive AI. In International Conference on Artificial General Intelligence,
pages 232–242, 2011.
32. Joscha Bach. Modeling motivation in MicroPsi 2. In International Conference on Artificial General Intelligence, pages
3–13, 2015.
33. Joscha Bach, Colin Bauer, and Ronnie Vuine. MicroPsi: Contributions to a broad architecture of cognition. In Annual
Conference on Artificial Intelligence, pages 7–18, 2007.
34. Pilar Bachiller, Pablo Bustos, and Luis J. Manso. Attentional Selection for Action in Mobile Robots. In Advances in
Robotics, Automation and Control, pages 111–136. 2008.
35. Alan D. Baddeley and Graham Hitch. Working Memory. Psychology of Learning and Motivation - Advances in Research
and Theory, 8(C):47–89, 1974.
36. Antonio Bandera, Juan Pedro Bandera, Pablo Bustos, Luis V. Calderita, Fernando Fern, Raquel Fuentetaja, Fran-
cisco Javier Garc, Ana Iglesias, J. Luis, Rebeca Marfil, Carlos Pulido, Christian Reuther, Adrian Romero-Garces, and
Cristina Suarez. CLARC: a Robotic Architecture for Comprehensive Geriatric Assessment. In Proceedings of the WAF2016,
2016.
37. Antonio Bandera and Pablo Bustos. Toward the Development of Cognitive Robots. In International Workshop on
Brain-Inspired Computing, 2013.
38. Anthony Barnes and Robert J. Hammell. Determining Information Technology Project Status using Recognition-primed
Decision- making enabled Collaborative Agents for Simulating Teamwork (R-CAST). In Proceedings of the Conference
on Information Systems Applied Research (CONISAR), 2008.
39. Burcu Arslan Barslanrugnl, Niels Taatgen Nataatgenrugnl, and Rineke Verbrugge. Modeling Developmental Transitions
in Reasoning about False Beliefs of Others. 2004.
40. Thomas M. Bartol, Cailey Bromer, Justin P. Kinney, Micheal A. Chirillo, Jennifer N. Bourne, Kristen M. Harris, and
Terrence J. Sejnowski. Nanoconnectomic Upper Bound on the Variability of Synaptic Plasticity. eLife, 4(210778), 2015.
41. F. Bellas, J. A. Becerra, and R. J. Duro. Induced Behavior in a Real Agent Using the Multilevel Darwinist Brain. In
International Work-Conference on the Interplay Between Natural and Artificial Computation, pages 425–434, 2005.
42. F. Bellas, J. A. Becerra, and R. J. Duro. Some experimental results with a two level memory management system in the
Multilevel Darwinist Brain. In Proceedings of the European Symposium on Artificial Neural Networks, Computational
Intelligence and Machine Learning, 2006.
43. F. Bellas and R. J. Duro. Some thoughts on the use of sampled fitness functions for the multilevel Darwinist brain.
Information Sciences, 161(3-4):159–179, 2004.
44. Francisco Bellas, Pilar Caamano, Andres Faina, and Richard J. Duro. Dynamic learning in cognitive robotics through a
procedural long term memory. Evolving Systems, 5(1):49–63, 2014.
45. Francisco Bellas, Richard J. Duro, Andrés Faiña, and Daniel Souto. Multilevel Darwinist Brain (MDB): Artificial evolution
in a cognitive architecture for real robots. IEEE Transactions on Autonomous Mental Development, 2(4):340–354, 2010.
46. Paul Bello, Will Bridewell, and Christina Wasylyshyn. Attentive and Pre-Attentive Processes in Multiple Object Tracking:
A Computational Investigation Modeling Object Construction and Tracking. In Proceedings of the 38th Annual Meeting
of the Cognitive Science Society, 2016.
47. D. P. Benjamin and Damian Lyons. A cognitive approach to classifying perceived behaviors. In Proc. SPIE 7710,
Multisensor, Multisource Information Fusion: Architectures, Algorithms, and Applications, volume 7710, 2010.
48. D. Paul Benjamin, Christopher Funk, and Damian Lyons. A cognitive approach to vision for a mobile robot. SPIE
Defense, Security, and Sensing, 2013.
49. D. Paul Benjamin, Damian Lyons, and Deryle Lonsdale. ADAPT: A Cognitive Architecture for Robotics. Proceedings of
the Sixth International Conference on Cognitive Modeling, (October):337–338, 2004.
50. Michal Bida, Martin Cerny, Jakub Gemrot, and Cyril Brom. Evolution of GameBots project. In International Conference
on Entertainment Computing, pages 397–400, 2012.
51. Cristina Boicu, Gheorghe Tecuci, and Mihai Boicu. Improving Agent Learning through Rule Analysis. In Proceedings of
the International Conference on Artificial Intelligence, 2005.
52. Mihai Boicu, Dorin Marcu, Cristina Boicu, and Bogdan Stanescu. Mixed-initiative Control for Teaching and Learning in
Disciple. In Proceedings of the IJCAI-03 Workshop on Mixed-Initiative Intelligent Systems, 2003.
54 Iuliia Kotseruba, John K. Tsotsos

53. Ladislau Bölöni. An Investigation into the Utility of Episodic Memory for Cognitive Architectures. In AAAI fall sympo-
sium: Advances in Cognitive Systems, 2012.
54. Jonathan P. Bona. MGLAIR: A Multimodal Cognitive Agent Architecture. PhD Thesis, 2013.
55. R. Peter Bonasso, R. James Firby, Erann Gat, David Kortenkamp, David P. Miller, and Marc G. Slack. Experiences with
an architecture for intelligent, reactive agents. Journal of Experimental & Theoretical Artificial Intelligence, (2-3):187–202,
1997.
56. R. Peter Bonasso and David Kortenkamp. Using a Layered Control Architecture to Alleviate Planning with Incomplete
Information. In Proceedings of the AAAI Spring Symposium\Planning with Incomplete Information for Robot Problems,
1996.
57. Daniel Borrajo, Anna Roubı́čková, and Ivan Serina. Progress in Case-Based Planning. ACM Computing Surveys, 47(2),
2015.
58. Roger Bostelman, Tsai Hong, Tommy Chang, William Shackleford, and Michael Shneier. Unstructured facility navigation
by applying the NIST 4D/RCS architecture. In Proceedings of International Conference on Cybernetics and Information
Technologies, Systems and Applications, pages 328–333, 2006.
59. Roger V. Bostelman, Adam Jacoff, and Robert Bunch. Delivery of an Advanced Double-Hull Ship Welding. In Third
International ICSC (International Computer Science Conventions) Symposia on Intelligent Industrial Automation and
Soft Computing, 1999.
60. C. Breazeal, A. Edsinger, P. Fitzpatrick, and B. Scassellati. Active vision for sociable robots. IEEE Transactions on
Systems, Man, and Cybernetics Part A: Systems and Humans., 31(5):443–453, 2001.
61. Cynthia Breazeal. Emotion and sociable humanoid robots. International Journal of Human Computer Studies, 59(1-
2):119–155, 2003.
62. Cynthia Breazeal. Toward sociable robots. Robotics and Autonomous Systems, 42(3-4):167–175, 2003.
63. Cynthia Breazeal and Lijin Aryananda. Recognition of affective communicative intent in robot-directed speech. Au-
tonomous Robots, 12(1):83–104, 2002.
64. Cynthia Breazeal and Rodney Brooks. Robot Emotion: A Functional Perspective. In Who Needs Emotions?: The Brain
Meets the Robot. Oxford University Press, 2004.
65. Cynthia Breazeal, Aaron Edsinger, Paul Fitzpatrick, and Brian Scassellati. Social Constraints on Animate Vision. In
Proceedings of the HUMANOIDS, 2000.
66. Cynthia Breazeal and Paul Fitzpatrick. That Certain Look: Social Amplification of Animate Vision. In Proceedings of
AAAI 2000 Fall Symposium, pages 18–22, 2000.
67. Cynthia Breazeal and Brian Scassellati. A context-dependent attention system for a social robot. In IJCAI International
Joint Conference on Artificial Intelligence, volume 2, pages 1146–1151, 1999.
68. Cynthia Breazeal and Brian Scassellati. Challenges in Building Robots That Imitate People. In K. Dautenhahn and
C. Nehaniv, editors, Imitation in Animals and Artifacts, number 1998, pages 363–390. MIT Press, Cambridge, MA, 2002.
69. John L. Bresina and Mark Drummond. Integrating planning and reaction: A preliminary report. In Proceedings of AAAI
Spring Symposium on Planning in Uncertain, Unpredictable, or Changing Environments. NASA Ames Research Center,
1990.
70. Bryan E Brett, Jeffrey A Doyal, David A Malek, Edward A Martin, David G Hoagland, and Martin N Anesgart. The
Combat Automation Requirements Testbed (CART) Task 5 Interim Report: Modeling A Strike Fighter Pilot Conducting
a Time Critical target Mission. Technical Report, 2002.
71. Timothy Brick, Paul Schermerhorn, and Matthias Scheutz. Speech and action: Integration of action and language for
mobile robots. In Proceedings of IEEE International Conference on Intelligent Robots and Systems, pages 1423–1428,
2007.
72. Will Bridewell and Paul F. Bello. Incremental Object Perception in an Attention-Driven Cognitive Architecture. Proceed-
ings of the 37th Annual Meeting of the Cognitive Science Society, pages 279–284, 2015.
73. Will Bridewell and Paul F. Bello. Inattentional Blindness in a Coupled Perceptual-Cognitive System. In Proceedings of
the 38th Annual Meeting of the Cognitive Science Society, 2016.
74. Cyril Brom, Klára Pešková, and Jiri Lukavsky. What Does Your Actor Remember? Towards Characters with a Full
Episodic Memory. In International Conference on Virtual Storytelling, pages 89–101, 2007.
75. Rodney A. Brooks. Planning is just a way of avoiding figuring out what to do next. Technical Report Working Paper
303, 1987.
76. Rodney A. Brooks. A robot that walks; emergent behaviors from a carefully evolved network. Neural computation,
1(2):253–262, 1989.
77. Rodney A. Brooks. Elephants don’t play chess. Robotics and Autonomous Systems, 6(1-2):3–15, 1990.
78. Rodney A. Brooks and Anita M. Flynn. Robot Beings. In Proceedings of the IEEE/RSJ International Workshop on
Intelligent Robots and Systems, pages 2–10, 1989.
79. N. Bruce and J. Tsotsos. Attention based on information maximization. In Journal of Vision, volume 7, pages 950–950,
2007.
A Review of 40 Years in Cognitive Architecture Research Core Cognitive Abilities and Practical Applications 55

80. P. Bustos, J. Martinez-Gomez, I. Garcia-Varea, L. Rodriguez-Ruiz, P. Bachiller, L. Calderita, L. J. Manso, A. Sanchez,

A. Bandera, and J. P. Bandera. Multimodal Interaction with Loki. In Workshop of Physical Agents, 2013.
81. Pablo Bustos, Luis J. Manso, Juan P. Bandera, Adrián Romero-Garcés, Luis V. Calderita, Rebeca Marfil, and Antonio
Bandera. A unified internal representation of the outer world for social robotics. In Proceedings of the Second Iberian
Robotics Conference, 2016.
82. Ayesha Javed Butt, Naveed Anwer Butt, Aneela Mazhar, Zafar Khattak, and Javaid Anjum Sheikh. The Soar of cognitive
architectures. In Proceedings of the International Conference on Current Trends in Information Technology, 2013.
83. Fergal Byrne. Symphony from Synapses: Neocortex as a Universal Dynamical Systems Modeller using Hierarchical Tem-
poral Memory. arXiv preprint arXiv:1512.05245, 2015.
84. Shi Cao, Yulin Qin, Lei Zhao, and Mowei Shen. Modeling the development of vehicle lateral control skills in a cognitive
architecture. Transportation Research Part F: Traffic Psychology and Behaviour, 32:1–10, 2015.
85. Jaime G. Carbonell, Jim Blythe, Oren Etzioni, Yolanda Gil, Robert Joseph, Dan Kahn, Craig Knoblock, Steven Minton,
P. Alicia, Scott Reilly, Manuela Veloso, and Xuemei Wang. PRODIGY4.0: TheManual and Tutorial. Technical Report
CMU-CS-92-150, 1992.
86. Gail A. Carpenter. Neural-network models of learning and memory: Leading questions and an emerging framework.
TRENDS in Cognitive Sciences, 5(3):114–118, 2001.
87. Gail A. Carpenter and Sai Chaitanya Gaddam. Biased ART: A neural architecture that shifts attention toward previously
disregarded features following an incorrect prediction. Neural Networks, 23(3):435–451, 2010.
88. Gail A. Carpenter and Stephen Grossberg. Neural dynamics of category learning and recognition: attention, memory
consolidation, and amnesia. Advances in Psychology, 42:233–290, 1987.
89. Gail A. Carpenter, Stephen Grossberg, and John H. Reynolds. ARTMAP: Supervised real-time learning and classification
of nonstationary data by a self-organizing neural network. Neural Networks, 4(5):565–588, 1991.
90. Gail A Carpenter, Stephen Grossberg, Adaptive Systems, and Neural Systems. Adaptive Resonance Theory. Encyclopedia
of Machine Learning and Data Mining, 2016.
91. Gail A. Carpenter and Boriana L. Milenova. ART Neural Networks for Medical Data Analysis and Fast Distributed
Learning. In Artificial neural networks in medicine and biology. Springer, London, 2000.
92. John B. Carroll. Human cognitive abilities: A survey of factor-analytic studies. Cambridge University Press, 1993.
93. Justine Cassell and Kristinn R. Thorisson. The power of a nod and a glance: Envelope vs. emotional feedback in animated
conversational agents. Applied Artificial Intelligence, 13(4-5):519–538, 1999.
94. Nicholas L. Cassimatis. Harnessing Multiple Representations for Autonomous Full-Spectrum Political, Military, Economic,
Social, Information and Infrastructure (PMESII) Reasoning. Final Technical Report AFRL-IF-RS-TR-2007-131, 2007.
95. Nicholas L. Cassimatis, J. Gregory Trafton, Magdalena D. Bugajska, and Alan C. Schultz. Integrating cognition, perception
and action through mental simulation in robots. Robotics and Autonomous Systems, 49(1-2):13–23, 2004.
96. Belkacem Chikhaoui, Helene Pigot, Mathieu Beaudoin, Guillaume Pratte, Philippe Bellefeuille, and Fernando Laudares.
Learning a Song: an ACT-R Model. Proceedings of the International Conference on Computational Intelligence, pages
405–410, 2009.
97. Henrik I. Christensen, Aaron Sloman, Geert-Jan Kruijff, and Jeremy L. Wyatt, editors. Cognitive Systems. 2010.
98. Carlo Ciliberto, Fabrizio Smeraldi, Lorenzo Natale, and Giorgio Metta. Online multiple instance learning applied to hand
detection in a humanoid robot. In Proceedings of IEEE International Conference on Intelligent Robots and Systems,
pages 1526–1532, 2011.
99. Robert E. Cochran, Frank J. Lee, and Eric Chown. Modeling Emotion: Arousal’s Impact on Memory. In Proceedings of
the 28th Annual Conference of the Cognitive Science Society, pages 1133–1138, 2006.
100. Matthew Conforth and Yan Meng. CHARISMA: A Context Hierarchy-based cognitive architecture for self-motivated
social agents. Proceedings of the International Joint Conference on Neural Networks, pages 1894–1901, 2011.
101. Matthew Conforth and Yan Meng. Self-reorganizing knowledge representation for autonomous learning in social agents.
Proceedings of the International Joint Conference on Neural Networks, pages 1880–1887, 2011.
102. Matthew Conforth and Yan Meng. Embodied Intelligent Agents with Cognitive Conscious and Unconscious Reasoning.
In Proceedings of the BMI ICBM, pages 15–20, 2012.
103. D. Coombs, K. Murphy, A. Lacaze, and S. Legowik. Driving Autonomously Offroad up to 35 Km/h. In Proceedings of
the IEEE Intelligent Vehicles Symposium, number Mi, pages 186–191, 2000.
104. Kevin Corker, Greg Pisanich, and Marilyn Bunzo. A cognitive system model for human/automation dynamics in airspace
management. In Proceedings of the First European/US Symposium on Air Traffic Management, 1997.
105. Nelson Cowan. Chapter 20 What are the differences between long-term, short-term, and working memory? In Progress
in Brain Research, volume 169, pages 323–338. Elsevier, 2008.
106. L. Andrew Coward. Modelling Memory and Learning Consistently from Psychology to Physiology. In Vassilis Cutsuridis,
Amir Hussain, and John G. Taylor, editors, Perception-Action Cycle, pages 63–133. 2011.
107. Michael T Cox. Metacognition in Computation: A selected research review. Artif. Intell., 169(2):104–141, 2005.
108. Michael T. Cox. Perpetual Self-Aware Cognitive Agents. AI Mag., 28(1):32, 2007.
56 Iuliia Kotseruba, John K. Tsotsos

109. Michael T. Cox. MIDCA: A Metacognitive, Integrated Dual-Cycle Architecture for Self-Regulated Autonomy. Computer
Science Technical Report No. CS-TR-5025, 2013.
110. Michael T. Cox, Tim Oates, Matthew Paisner, and Donald Perlis. Noting Anomalies in Streams of Symbolic Predicates
Using A-Distance. Advances in Cognitive Systems, 2:167–184, 2012.
111. Eric Crawford, Matthew Gingerich, and Chris Eliasmith. Biologically Plausible, Human-scale Knowledge Representation.
Cognitive Science, 40(4):412–417, 2015.
112. Daniel Cuadrado and Yago Saez. Chuck Norris rocks! In Proceedings of the IEEE Symposium on Computational Intelli-
gence and Games, pages 69–74, 2009.
113. Yuwei Cui, Subutai Ahmad, and Jeff Hawkins. Continuous online sequence learning with an unsupervised neural network
model. arXiv preprint arXiv:1512.05463, 2015.
114. Jared F Danker and John R Anderson. The Ghosts of Brain States Past : Remembering Reactivates the Brain Regions
Engaged During Encoding. 136(1):87–102, 2010.
115. Dustin Dannenhauer, Michael T. Cox, Shubham Gupta, Matt Paisner, and Don Perlis. Toward meta-level control of
autonomous agents. In Proceedings of the 5th Annual International Conference on Biologically Inspired Cognitive Archi-
tectures, volume 41, pages 226–232. Elsevier Masson SAS, 2014.
116. M. Demetgul, I. N. Tansel, and S. Taskin. Fault diagnosis of pneumatic systems with artificial neural network algorithms.
Expert Systems with Applications, 36(7):10512–10519, 2009.
117. Nate Derbinsky and John E. Laird. Competence-Preserving Retention of Learned Knowledge in Soar’s Working and
Procedural Memories. In Proceedings of the 11th International Conference on Cognitive Modeling, 2012.
118. Nate Derbinsky and John E Laird. Computationally Efficient Forgetting via Base-Level Activation. In Proceedings of the
11th International Conference on Cognitive Modeling, pages 109–110, 2012.
119. Stephen Deutsch and Nichael Cramer. Omar Human Performance Modeling in a Decision Support Experiment. In
Proceedings of the Human Factors and Ergonomics Society 42nd Annual Meeting, pages 1232–1236, 1998.
120. Stephen Deutsch and Nichael Cramer. Omar Human Performance Modeling in a Decision Support Experiment. In
Proceedings of the Human Factors and Ergonomics Society 42nd Annual Meeting, pages 1232–1236, 1998.
121. Stephen E. Deutsch. UAV Operator Human Performance Models. Technical Report AFRL-RI-RS-TR-2006-0158, 2006.
122. Stephen E. Deutsch, Jean Macmillan, Nichael L. Camer, and Sonu Chopra. Operability Model Architecture: Demonstration
Final Report. Technical Report AL/HR-TR-1996-0161, 1997.
123. Alessandro Di Nuovo, Vivian M. De La Cruz, and Angelo Cangelosi. Grounding fingers, words and numbers in a cognitive
developmental robot. In Proceedings of the IEEE Symposium on Computational Intelligence, Cognitive Algorithms, Mind,
and Brain, 2014.
124. Mark D’Inverno, Michael Luck, Michael Georgeff, David Kinny, and Michael Wooldridge. The dMARS architecture: A
specification of the distributed multi-agent reasoning system. Autonomous Agents and Multi-Agent Systems, 9(1-2):5–53,
2004.
125. Cosimo Distante, Pietro Siciliano, and Lorenzo Vasanelli. Odor discrimination using adaptive resonance theory. Sensors
and Actuators, 69(3):248–252, 2000.
126. S.A. Douglass, J. Ball, and S. Rodgers. Large Declarative Memories in ACT-R. Proceedings of the 9th international
conference on cognitive modeling, page 234, 2009.
127. Mark Drummond and John Bresina. Planning for control. In Proceedings of the 5th IEEE International Symposium on
Intelligent Control, pages 657–662, 1990.
128. W. Duch, Richard J. Oentaryo, and Michel Pasquier. Cognitive Architectures: Where do we go from here? In Pei Wang,
Ben Goertzel, and Stan Franklin, editors, Frontiers in Artificial Intelligence and Applications, volume 171, pages 122–136.
IOS Press, 2008.
129. Wlodzislaw Duch. Computational Creativity. In IEEE World Congr. Comput. Intell., 2006.
130. Wlodzislaw Duch. Towards comprehensive foundations of computational intelligence. Studies in Computational Intelli-
gence, 63:261–316, 2007.
131. R. J. Duro, F. Bellas, and J. A. Becerra. Evolutionary Architecture for Lifelong Learning and Real-Time Operation in
Autonomous Robots. In P. Angelov, D. P. Filev, and N. Kasabov, editors, Evolving Intelligent Systems: Methodology and
Applications, pages 365–400. 2010.
132. Gerald M. Edelman. Learning in and from brain-based devices. Science, 318(5853):1103–1105, 2007.
133. Chris Eliasmith and Carter Kolbeck. Marr’s Attacks: On Reductionism and Vagueness. Topics in Cognitive Science,
pages 1–13, 2015.
134. Chris Eliasmith and Terrence C Stewart. A Large-Scale Model of the Functioning Brain. Science, 338(1202), 2012.
135. Les Elkins, Drew Sellers, and W. Reynolds Monach. The Autonomous Maritime Navigation (AMN) Project: Field Tests,
Autonomous and Cooperative Behaviors, Data Fusion, Sensors, and Vehicles. Journal of Field Robotics, 27(6):81–86,
2010.
136. S. L. Epstein. Metaknowledge for Autonomous Systems. In Proceedings of AAAI Spring Symposium on Knowledge
Representation and Ontology for Autonomous Systems, 2004.
A Review of 40 Years in Cognitive Architecture Research Core Cognitive Abilities and Practical Applications 57

137. Susan L. Epstein. The role of memory and concepts in learning. Minds and Machines, 2(3):239–265, 1992.
138. Susan L. Epstein. For the right reasons: The FORR architecture for learning in a skill domain. Cognitive Science,
18(3):479–511, 1994.
139. Susan L. Epstein. On Heuristic Reasoning, Reactivity, and Search. In Proceedings of IJCAI-95, pages 454–461, 1995.
140. Susan L. Epstein. Learning to Play Expertly: A Tutorial on Hoyle. In Johannes Furnkranz and Miroslav Kubat, editors,
Machines that learn to play games, pages 153–178. Nova Science Publishers, 2001.
141. Susan L. Epstein, Eugene C. Freuder, Richard Wallace, Anton Morozov, and Bruce Samuels. The Adaptive Constraint
Engine. In Proceedings of the International Conference on Principles and Practice of Constraint Programming, pages
525–540, 2002.
142. Susan L. Epstein, Rebecca Passonneau, Joshua Gordon, and Tiziana Ligorio. The Role of Knowledge and Certainty in
Understanding for Dialogue. In AAAI Fall Symposium: Advances in Cognitive Systems, 2012.
143. Susan L Epstein and Smiljana Petrovic. Learning Expertise with Bounded Rationality and Self-awareness. In Metarea-
soning: Thinking about Thinking. MIT Press Scholarship Online, 2008.
144. Oren Etzioni. Acquiring search-control knowledge via static analysis. Artificial Intelligence, 62(2):255–301, 1993.
145. Rick Evertsz, Matteo Pedrotti, Paolo Busetta, Hasan Acar, and Frank E. Ritter. Populating VBS2 with Realistic Virtual
Actors. In Proceedings of the 18th Conference on Behavior Representation in Modeling and Simulation, 2009.
146. Rick Evertsz, F. E. Ritter, Simon Russell, and David Shepherdson. Modeling rules of engagement in computer generated
forces. Proceedings of the 16th Conference on Behavior Representation in Modeling and Simulation, pages 123–34, 2007.
147. Usef Faghihi. The use of emotions in the implementation of various types of learning in a cognitive agent. PhD Thesis,
2011.
148. Usef Faghihi, Philippe Fournier-Viger, and Roger Nkambou. Implementing an efficient causal learning mechanism in a cog-
nitive tutoring agent. International Conference on Industrial, Engineering and Other Applications of Applied Intelligent
Systems, 2011.
149. Usef Faghihi, Philippe Fournier-Viger, and Roger Nkambou. CELTS: A cognitive tutoring agent with human-like learning
capabilities and emotions. In Smart Innovation, Systems and Technologies. Springer Berlin Heidelberg, 2013.
150. Usef Faghihi and Stan Franklin. The LIDA Model as a Foundational Architecture for AGI. Theoretical Foundations of
Artificial General Intelligence, 4:103–121, 2012.
151. Usef Faghihi, Pierre Poirier, and Othalia Larue. Emotional Cognitive Architectures. In International Conference on
Affective Computing and Intelligent Interaction, 2011.
152. X. Fan, B. Sun, S. Sun, M. McNeese, and J. Yen. RPD-enabled agents teaming with humans for multi-context decision
making. In Proceedings of the International Conference on Autonomous Agents, 2006.
153. Xiaocong Fan, Michael McNeese, Bingjun Sun, Timothy Hanratty, Laurel Allender, and John Yen. Human-agent collab-
oration for time-stressed multicontext decision making. IEEE Transactions on Systems, Man, and Cybernetics Part A:
Systems and Humans, 40(2):306–320, 2010.
154. Xiaocong Fan, Michael McNeese, and John Yen. NDM-Based Cognitive Agents for Supporting Decision-Making Teams.
Human-Computer Interaction, 25(3):195–234, 2010.
155. Arash Fazl, Stephen Grossberg, and Ennio Mingolla. View-invariant object category learning, recognition, and search:
How spatial and object attention are coordinated using surface-based attentional shrouds. Cognitive Psychology, 58(1),
2009.
156. Anna Fedor, István Zachar, András Szilágyi, and Michael Öllinger. Cognitive Architecture with Evolutionary Dynamics
Solves Insight Problem. Frontiers in Psychology, 8:1–15, 2017.
157. Veselina Feldman and Boicho Kokinov. Anxiety restricts the analogical search in an analogy generation task. New Frontiers
in Analogy Research, 2009.
158. Eugene Fink and Jim Blythe. Prodigy bidirectional planning. Journal of Experimental & Theoretical Artificial Intelligence,
17(3):161–200, 2005.
159. James R. Firby, Roger E. Kahn, Peter N. Prokopowicz, and Michael J. Swain. An Architecture for Vision and Action. In
Proceedings of the 14th international joint conference on Artificial intelligence, 1995.
160. Robert James Firby. Adaptive Execution in Complex Dynamic Worlds. PhD thesis, 1989.
161. John H Flavell. Metacognition and Cognitive Monitoring: A New Area of Cognitive-Developmental Inquiry. Am. Psychol.,
34(10):906–911, 1979.
162. Jason G. Fleischer and Gerald M. Edelman. Brain-Based Devices: An embodied approach to linking nervous system
structure and function to behavior. IEEE Robotics and Automation Magazine, 16(3):33–41, 2009.
163. Jason G. Fleischer and Jeffrey L. Krichmar. Sensory Integration and Remapping in a Model of the Medial Temporal Lobe
During Maze Navigaion by a Brain-Based Device. Journal of Integrative Neuroscience, 6(3):403–431, 2007.
164. A. M. Flynn, R. A. Brooks, W. M. Wells, and D. S. Barrett. The world’s largest one cubic inch robot. In Proceedings of
IEEE Conference on Microelectromechanical Systems, pages 98–101, 1989.
165. Kenneth D Forbus, Ronald W Ferguson, and Andrew Lovett. Extending SME to Handle Large-Scale Cognitive Modeling.
Cognitive Science, pages 1–50, 2016.
58 Iuliia Kotseruba, John K. Tsotsos

166. Kenneth D. Forbus, Matthew Klenk, and Thomas Hinrichs. Companion Cognitive Systems: Design Goals and Lessons
Learned. IEEE Intelligent Systems, PP(99):36–46, 2009.
167. Douglas Foxvog. Cyc. In Theory and Applications of Ontology: Computer Applications, pages 259–278. 2010.
168. S. Franklin. Modeling Consciousness and Cognition in Software Agents. In Proceedings of the Third International
Conference on Cognitive Modeling, pages 27–58, 2000.
169. Stan Franklin. Learning in ”Conscious” Software Agents. In Workshop on Development and Learning, 2000.
170. Stan Franklin. A Foundational Architecture for Artificial General Intelligence. Advances in Artificial General Intelligence:
Concepts, Architectures and Algorithms, pages 36–54, 2007.
171. Stan Franklin, Tamas Madl, Steve Strain, Usef Faghihi, Daqi Dong, Sean Kugele, Javier Snaider, Pulin Agrawal, and
Sheng Chen. A LIDA cognitive model tutorial. Biologically Inspired Cognitive Architectures, 16, 2016.
172. Stan Franklin, Steve Strain, Javier Snaider, Ryan McCall, and Usef Faghihi. Global Workspace Theory, its LIDA model
and the underlying neuroscience. Biologically Inspired Cognitive Architectures, 1:32–43, 2012.
173. M. Freed and R. Remington. Making human-machine system simulation a practical engineering tool: An APEX overview.
In Proceedings of the 3rd International Conference on Cognitive Modelling, 2000.
174. M. A. Freed. Simulating human performance in complex, dynamic environments. PhD Thesis, (June), 1998.
175. Scott E. Friedman and Kenneth D. Forbus. An Integrated Systems Approach to Explanation-Based Conceptual Change.
Association for the Advancement of Artificial Intelligence, 2010.
176. Scott E. Friedman, Kenneth D. Forbus, and Bruce Sherin. Constructing & revising commonsense science explanations: A
metareasoning approach. AAAI Fall Symposium on Advances in Cognitive Systems, 2011.
177. Simone Frintrop, Erich Rome, and Henrik I. Christensen. Computational Visual Attention Systems and their Cognitive
Foundations: A Survey. ACM Transactions on Applied Perception, 7(1), 2010.
178. Jeffrey From, Patrick Perrin, Daniel O’Neill, and John Yen. Supporting the Commander’s information requirements:
Automated support for battle drill processes using R-CAST. In Proceedings of the IEEE Military Communications
Conference MILCOM, 2011.
179. Erann Gat. Integrating Planning and Reacting in a Heterogeneous Asynchronous Architecture for Controlling Real-World
Mobile Robots. AAAI, pages 809–815, 1992.
180. Erann Gat and Greg Dorais. Robot navigation by conditional sequencing. In Proceedings of the International Conference
on Robotics and Automation, pages 1293–1299, 1994.
181. Jakub Gemrot, Cyril Brom, Rudolf Kadlec, Michal Bida, Ondrej Burkert, Michal Zemčák, Radek Pı́bil, and Tomáš Plch.
Pogamut 3 – Virtual Humans Made Simple. Advances in Cognitive Science, pages 211–243, 2010.
182. Jakub Gemrot, Rudolf Kadlec, Michal Bida, Ondrej Burkert, Radek Pibil, Jan Havlicek, Lukas Zemcak, Juraj Simlovic,
Radim Vansa, Michal Stolba, Tomas Plch, and Cyril Brom. Pogamut 3 can assist developers in building AI (not only) for
their videogame agents. Agents for games and simulations, pages 1–15, 2009.
183. Michael Georgeff, Barney Pell, Martha Pollack, Milind Tambe, and Michael Wooldridge. The Belief-Desire-Intention Model
of Agency. In International Workshop on Agent Theories, Architectures, and Languages, 1998.
184. Michael P. Georgeff and Francois Felix Ingrand. Decision-Making in an Embedded Reasoning System. In Proceedings of
the Eleventh International Joint Conference on Artificial Intelligence (IJCAI-89), 1989.
185. Michael P. Georgeff and Amy L. Lansky. Procedural Knowledge. In Proceedings of the IEEE, volume 74, pages 1383–1398,
1986.
186. F. Gobet and P. C. Lane. Chunking Mechanisms and Learning. In N. M. Seel, editor, Encyclopedia of the Sciences of
Learning, pages 541–544. Springer, New York, NY, 2012.
187. Fernand R. Gobet. Memory for the meaningless: How chunks help. In Proceedings of the 20th Meeting of the Cognitive
Science Society, pages 398–403, 2008.
188. Ben Goertzel. A pragmatic path toward endowing virtually-embodied AIs with human-level linguistic capability. In
Proceedings of the International Joint Conference on Neural Networks, 2008.
189. Ben Goertzel. Perception processing for general intelligence: Bridging the symbolic/subsymbolic gap. AI, 2012.
190. Ben Goertzel, Hugo De Garis, Cassio Pennachin, Nil Geisweiller, Samir Araujo, Joel Pitt, Shuo Chen, Ruiting Lian, Min
Jiang, Ye Yang, and Deheng Huang. OpenCogBot: Achieving Generally Intelligent Virtual Agent Control and Humanoid
Robotics via Cognitive Synergy. In Proceedings of International Conference on Artifical Intelligence, 2010.
191. Ben Goertzel, Ruiting Lian, Itamar Arel, Hugo de Garis, and Shuo Chen. A world survey of artificial brain projects, Part
II: Biologically inspired cognitive architectures. Neurocomputing, 74(1-3):30–49, 2010.
192. Ben Goertzel and Cassio Pennachin. The Novamente Artificial Intelligence Engine. In Artificial General Intelligence,
pages 63–129. Springer Berlin Heidelberg, 2007.
193. Ben Goertzel, Cassio Pennachin, Nil Geissweiller, Moshe Looks, Andre Senna, Welter Silva, Ari Heljakka, and Carlos
Lopes. An Integrative Methodology for Teaching Embodied Non-Linguistic Agents, Applied to Virtual Animals in Second
Life. Frontiers in Artificial Intelligence and Applications, 171:161–175, 2008.
194. Ben Goertzel, Cassio Pennachin, and Nil Geisweiller. Brief Survey of Cognitive Architectures. In Engineering General
Intelligence, Part 1, pages 101–142. Atlantis Press, 2014.
A Review of 40 Years in Cognitive Architecture Research Core Cognitive Abilities and Practical Applications 59

195. Ben Goertzel, Cassio Pennachin, and SA De Souza. An inferential dynamics approach to personality and emotion driven
behavior determination for virtual animals. AISB 2008 Convention on Communication, Interaction and Social Intelli-
gence, 2008.
196. Ben Goertzel, Ted Sanders, and Jade O’Neill. Integrating deep learning based perception with probabilistic logic via
frequent pattern mining. In International Conference on Artificial General Intelligence, 2013.
197. Ben Goertzel and Gino Yu. A cognitive API and its application to AGI intelligence assessment. In International Conference
on Artificial General Intelligence, pages 242–245, 2014.
198. Joshua Gordon and Susan L. Epstein. Learning to balance grounding rationales for dialogue systems. In Proceedings of
SIGDIAL Conference, pages 266–271, 2011.
199. Brian F. Gore, Becky L. Hooey, Christopher D. Wickens, and Shelly Scott-Nash. A computational implementation of a
human attention guiding mechanism in MIDAS v5. In International Conference on Digital Human Modeling, 2009.
200. Jan Gosmann, Terrence C Stewart, and Thomas Wennekers. A Spiking Neuron Model of Word Associations for the
Remote Associates Test. Frontiers in Psychology, 8, 2017.
201. Stephen Grossberg. The Link between Brain Learning, Attention, and Consciousness. Consciousness and Cognition,
8:1–44, 1999.
202. Stephen Grossberg. Resonant Neural Dynamics of Speech Perception. Technical Report CAS/CNS-TR-02-008, 2003.
203. Stephen Grossberg. Towards a unified theory of neocortex: laminar cortical circuits for vision and cognition. Progress in
Brain Research, 165:79–104, 2007.
204. Stephen Grossberg, Krishna K. Govindarajan, Lonce L. Wyse, and Michael A. Cohen. ARTSTREAM: a neural network
model of auditory scene analysis and source segregation. Neural Networks, 17(4):511–536, 2004.
205. Stephen Grossberg and Christopher W. Myers. The Resonant Dynamics of Speech Perception: Interword Integration and
Duration-Dependent Backward Effects. Psychological Review, 107(4), 2015.
206. Glenn Gunzelmann, Joshua B. Gross, Kevin A. Gluck, and David F. Dinges. Sleep deprivation and sustained attention
performance: Integrating mathematical and cognitive modeling. Cognitive Science, 33(5):880–910, 2009.
207. Patrick Hammer, Tony Lofthouse, and Pei Wang. The OpenNARS implementation of the Non-Axiomatic Reasoning
System. In International Conference on Artificial General Intelligence, 2016.
208. Cosmo Harrigan, Ben Goertzel, Matthew Ikle, Amen Belayneh, and Gino Yu. Guiding probabilistic logical inference with
nonlinear dynamical attention allocation. In International Conference on Artificial General Intelligence, pages 238–241,
2014.
209. Sandra Hart, David Dahn, Adolph Atencio, and Michael K. Dalal. Evaluation and Application of MIDAS v2.0. SAE
Technical Paper 2001-01-2648, 2001.
210. Nick Hawes, Jeremy L. Wyatt, Mohan Sridharan, Marek Kopicki, Somboon Hongeng, Ian Calvert, Aaron Sloman, Geert-
Jan Kruijff, Henrik Jacobsson, Michael Brenner, Danijel Skocaj, Alen Vrecko, Nikodem Majer, and Michael Zillich. The
Playmate System. In Cognitive Systems. 2010.
211. Jeff Hawkins and Dileep George. Hierarchical temporal memory: Theory and applications, 2006.
212. Barbara Hayes-Roth. An architecture for adaptive intelligent systems. Artificial Intelligence, 72(1-2):329–365, 1995.
213. Barbara Hayes-Roth. A domain-specific software architecture for a class of intelligent patient monitoring agents. Journal
of Experimental & Theoretical Artificial Intelligence, 8(2), 1996.
214. Barbara Hayes-Roth and Robert Van Gent. Story-Making with Improvisational Puppets and Actors. In Proc. first Int.
Conf. Auton. agents., 1995.
215. Barbara Hayes-Roth and Robert Van Gent. Story-Making with Improvisational Puppets and Actors. In Proceedings of
the first international conference on Autonomous agents., 1995.
216. Barbara Hayes-Roth, Philippe Lalanda, Philippe Morignot, Karl Pfleger, and Marko Balabanovic. Plans and Behavior in
Intelligent Agents. KSL Report No. 93-43, 1993.
217. Barbara Hayes-Roth, Richard Washington, David Ash, Rattikorn Hewett, Anne Collinot, Angel Vina, and Adam Seiver.
Guardian: A prototype intelligent agent for intensive-care monitoring. Artificial Intelligence In Medicine, 4(2):165–185,
1992.
218. Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Delving Deep into Rectifiers : Surpassing Human-Level
Performance on ImageNet Classification. In ICCV, 2014.
219. Sebastien Helie and Ron Sun. Incubation, Insight, and Creative Problem Solving: A Unified Theory and a Connectionist
Model. Psychol. Rev., 117(3), 2010.
220. Sebastien Helie and Ron Sun. An integrative account of memory and reasoning phenomena. New Ideas in Psychology,
35(1):36–52, 2014.
221. Ari Heljakka, Ben Goertzel, Welter Silva, Cassio Pennachin, Andre Senna, and Izabela Goertzel. Probabilistic Logic Based
Reinforcement Learning of Simple Embodied Behaviors in a 3D Simulation World. Frontiers in Artifical Intelligence and
Applications, 157:253–275, 2007.
222. T. C. Henderson, H. Peng, K. Sikorski, N. Deshpande, and E. Grant. The Cognitive Symmetry Engine: An Active Approach
to Knowledge. In Proceedings of the IROS 2011 workshop on Knowledge Representation for Autonomous Robots, 2011.
60 Iuliia Kotseruba, John K. Tsotsos

223. Thomas C Henderson and Anshul Joshi. The Cognitive Symmetry Engine. Technical Report UUCS-13-004, 2013.
224. Thomas C. Henderson, Anshul Joshi, and Edward Grant. From Sensorimotor Data to Concepts: The Role of Symmetry.
Technical Report UUCS-12-005, 2012.
225. Seth Herd, Andrew Szabados, Yury Vinokurov, Christian Lebiere, Ashley Cline, and Randall C. O’Reilly. Integrating
theories of motor sequencing in the SAL hybrid architecture. Biologically Inspired Cognitive Architectures, 8:98–106,
2014.
226. Seth A. Herd, Kai A. Krueger, Trenton E. Kriete, Tsung Ren Huang, Thomas E. Hazy, and Randall C. O’Reilly. Strategic
cognitive sequencing: A computational cognitive neuroscience approach. Computational Intelligence and Neuroscience,
2013, 2013.
227. Jose Hernandez-Orallo. Evaluation in artificial intelligence: From task-oriented to ability-oriented measurement. Artif.
Intell. Rev., 48(3):397–447, 2017.
228. E. Tory Higgins and Baruch Eitam. Priming. . . Shmiming: It’s About Knowing When & Why Stimulated Memory Rep-
resentations Become Active. Social Cognition, 32:1–33, 2014.
229. Melanie Hilario. An Overview of Strategies for Neurosymbolic Integration. In Ron Sun and Frederic Alexandre, editors,
Connectionist-symbolic integration: From unified to hybrid approaches, pages 13–35. Psychology Press, 1997.
230. Thomas R. Hinrichs and Kenneth D. Forbus. Analogical learning in a turn-based strategy game. In Proceedings of
International Joint Conference on Artificial Intelligence, pages 853–858, 2007.
231. Thomas R. Hinrichs and Kenneth D. Forbus. X Goes First: Teaching Simple Games through Multimodal Interaction.
Advances in Cognitive Systems, 3:218, 2014.
232. Douglas Hofstadter. How could a COPYCAT ever be creative? AAAI Technical Report SS-93-01, pages 8–21, 1993.
233. Clay B. Holroyd and Michael G. H. Coles. The Neural Basis of Human Error Processing: Reinforcement Learning,
Dopamine, and the Error-Related Negativity. Psychological Review, 109(4):679–709, 2002.
234. Tsai Hong Hong, Stephen B. Balakirsky, Elena Messina, Tommy Chang, and Michael Shneier. A Hierarchical World Model
for an Autonomous Scout Vehicle. In 16th Annual International Symposium On Aerospace/Defense Sensing, Simulation,
And Controls (SPIE 2002), pages 343–354, 2002.
235. Becky L. Hooey, Brian F. Gore, Christopher D. Wickens, Shelly Scott-Nash, Connie M. Socash, Ellen Salud, and David C.
Foyle. Human Modelling in Assisted Transportation. In Proceeding of the Human Modeling in Assisted Transportation
Conference, pages 327–333, 2010.
236. Xiao Huang and Juyang Weng. Inherent Value Systems for Autonomous Mental Development. International Journal of
Humanoid Robotics, 4(2):407–433, 2007.
237. Eric Huber and David Kortenkamp. Using Stereo Vision to Pursue Moving Agents with a Mobile Robot. In Proceedings
of the IEEE International Conference on Robotics and Automation, 1995.
238. Eva Hudlicka. Modeling Affect Regulation and Induction. In Proceedings of the AAAI Fall Symposium 2001, ”Emotional
and Intelligent II: The Tangled Knot of Social Cognition”, 2001.
239. Eva Hudlicka. This time with feeling: Integrated model of trait and state effects on cognition and behavior. Applied
Artificial Intelligence, 16:611–641, 2002.
240. Eva Hudlicka. Beyond Cognition: Modeling Emotion in Cognitive Architectures. In Proceedings of the Sixth International
Conference on Cognitive Modeling, pages 118–123, 2004.
241. Eva Hudlicka. A computational model of emotion and personality: Applications to psychotherapy research and practice.
In Proceedings of the 10th Annual CyberTherapy Conference: A Decade of Virtual Reality, 2005.
242. Eva Hudlicka. Modeling Effects of Emotion and Personality on Political Decision-Making. In Programming for Peace,
pages 355–411. Springer, 2006.
243. Eva Hudlicka. Modeling the mechanisms of emotion effects on cognition. In Proceedings of the AAAI Fall Symposium on
Biologically Inspired Cognitive Architectures, pages 82–86, 2008.
244. Eva Hudlicka. Challenges in Developing Computational Models of Emotion and Consciousness. International Journal of
Machine Consciousness, 1(1):131–153, 2009.
245. Eva Hudlicka. Modeling Cultural and Personality Biases in Decision Making. In Proceedings of the 3rd International
Conference on Applied Human Factors and Ergonomics (AHFE), 2010.
246. Eva Hudlicka. Computational Analytical Framework for Affective Modeling: Towards Guidelines for Designing. In Psy-
chology and Mental Health: Concepts, Methodologies, Tools, and Applications: Concepts, Methodologies, Tools, and Ap-
plications, pages 1–64. 2016.
247. Eva Hudlicka and Gerald Matthews. Affect, Risk and Uncertainty in Decision-Making. An Integrated Computational-
Empirical Approach. Final Report, 2009.
248. Eva Hudlicka, Greg Zacharias, and Joseph Psotka. Increasing realism of human agents by modeling individual differences:
Methodology, architecture, and testbed. Simulating Human Agents, American Association for Artificial Intelligence Fall
2000 Symposium Series, pages 53–59, 2000.
249. Terry Huntsberger. Cognitive architecture for mixed human-machine team interactions for space exploration. In IEEE
Aerospace Conference Proceedings, 2011.
A Review of 40 Years in Cognitive Architecture Research Core Cognitive Abilities and Practical Applications 61

250. Terry Huntsberger, Hrand Aghazarian, Andrew Howard, and David C. Trotz. Stereo vision-based navigation for au-
tonomous surface vessels. Journal of Field Robotics, 28(1):3–18, 2011.
251. Terry Huntsberger and Adrian Stoica. Envisioning Cognitive Robots for Future Space Exploration. SPIE Defense,
Security, and Sensing, 2010.
252. Terry Huntsberger and Gail Woodward. Intelligent Autonomy for Unmanned Surface and Underwater Vehicles. In
Proceedings of the OCEANS’11, pages 1–10, 2011.
253. Ryutaro Ichise. An Analysis of the CHC model for Comparing Cognitive Architectures. Procedia Comput. Sci., 88:239–244,
2016.
254. Matthew Ikle and Ben Goertzel. Nonlinear-dynamical attention allocation via information geometry. In International
Conference on Artificial General Intelligence, 2011.
255. Laurent Itti, Christof Koch, and Ernst Niebur. A Model of Saliency-Based Visual Attention for Rapid Scene Analysis.
IEEE Transactions on Pattern Analysis and Machine Intelligence, 20(11):1254–1259, 1998.
256. Serena Ivaldi, Natalia Lyubova, Damien Gerardeaux-Viret, Alain Droniou, Salvatore M. Anzalone, Mohamed Chetouani,
David Filliat, and Olivier Sigaud. Perception and human interaction for developmental learning of objects and affordances.
In IEEE-RAS International Conference on Humanoid Robots, pages 248–254, 2012.
257. Serena Ivaldi, Sao Mai Nguyen, Natalia Lyubova, Alain Droniou, Vincent Padois, David Filliat, Pierre Yves Oudeyer,
and Olivier Sigaud. Object learning through active exploration. EEE Transactions on Autonomous Mental Development,
6(1):56–72, 2014.
258. Marek Jaszuk and Janusz A Starzyk. Building Internal Scene Representation in Cognitive Agents. In Knowledge, Infor-
mation and Creativity Support Systems: Recent Trends, Advances and Solutions, pages 479–491. 2016.
259. R. Jensen and M. Veloso. Interleaving deliberative and reactive planning in dynamic multi-agent domains. In Proceedings
of the AAAI Fall Symposium on on Integrated Planning for Autonomous Agent Architectures, 1998.
260. David J. Jilk, Christian Lebiere, Randall C. O’Reily, and John R. Anderson. SAL: An explicitly pluralistic cognitive
architecture. Journal of Experimental and Theoretical Artificial Intelligence, 20(3):197–218, 2008.
261. Randolph M. Jones, John E. Laird, Paul E. Nielsen, Karen J. Coulter, Patrick Kenny, and Frank V. Koss. Automated
Intelligent Pilots for Combat Flight Simulation. AI Magazine, 20(1):27–42, 1999.
262. Randolph M. Jones, Robert E. III Wray, and Michael van Lent. Practical Evaluation of Integrated Cognitive Systems.
Adv. Cogn. Syst., 1:83–92, 2012.
263. Steven J. Jones, Arthur R. Wandzel, and John E. Laird. Efficient Computation of Spreading Activation Using Lazy
Evaluation. In Proceedings of the International Conference on Cognitive Modeling, 2016.
264. Gudny Ragna Jonsdottir and Kristinn R. Thórisson. A Distributed Architecture for Real-time Dialogue and On-task
Learning of Efficient Co-operative Turn-taking. In N. Campbell and R. Matej, editors, Coverbal Synchrony in Human-
Machine Interaction, pages 293–323. CRC Press, 2013.
265. Marcel Adam Just and Patricia A. Carpenter. A capacity theory of comprehension: Individual differences in working
memory. Psychological Review, 99(1):122–149, 1992.
266. Marcel Adam Just and Sashank Varma. The organization of thinking: What functional brain imaging reveals about the
neuroarchitecture of complex cognition. Cognitive, affective & behavioral neuroscience, 7(3):153–191, 2007.
267. Rudolf Kadlec, Jakub Gemrot, Michal Bida, Ondrej Burkert, Jan Havlicek, Lukas Zemcak, Radek Pibil, Radim Vansa,
and Cyril Brom. Extensions and applications of Pogamut 3 platform. In International Workshop on Intelligent Virtual
Agents, 2009.
268. K. Kawamura, M. Cambron, K. Fujiwara, and J. Barile. A Cooperative Robotic Aid System. In Proceedings of the
Conference on Virtual Reality Systems, Teleoperation and Beyond Speech Recognition, 1993.
269. Kazuhiko Kawamura, Stephen M. Gordon, Palis Ratanaswasd, Erdem Erdemir, and Joseph F. Hall. Implementation of
Cognitive Control for a Humanoid Robot. International Journal of Humanoid Robotics, 5(4):547–586, 2008.
270. Kazuhiko Kawamura, R. Alan II Peters, Robert E. Bodenheimer, Nilanjan Sarkar, Juyi Park, Charles A. Clifton, and
Albert W. Spratley. A Parallel Distributed Cognitive Control System for a Humanoid Robot. International Journal of
Humanoid Robotics, 1(1):65–93, 2004.
271. Assem Kaylani, Michael Georgiopoulos, Mansooreh Mollaghasemi, and Georgios C. Anagnostopoulos. AG-ART: An
adaptive approach to evolvong ART architectures. Neurocomputing, 72:2079–2092, 2009.
272. Smadar T. Kedar and Kathleen B. McKusick. There is No Free Lunch: Tradeoffs in the Utility of Learned Knowledge. In
Proceedings of the First International Conference on Artificial Intelligence Planning Systems, pages 281–282, 1992.
273. Troy D. Kelley. Symbolic and Sub-Symbolic Representations in Computational Models of Human Cognition: What Can
be Learned from Biology? Theory & Psychology, 13(6):847–860, 2003.
274. W. Kennedy and J. G. Trafton. Long-term symbolic learning in Soar and ACT-R. In Proceedings of the Seventh Inter-
national Conference on Cognitive Modelling, pages 166–171, 2006.
275. W. G. Kennedy and K. A. De Jong. Characteristics of Long-Term Learning in Soar and its Application to the Utility
Problem. Proceedings of the fifth international conference on machine learning, pages 337–344, 2003.
276. Bahador Khaleghi, Alaa Khamis, and Fakhreddine O. Karray. Multisensor data fusion: A review of the state-of-the-art.
Inf.ormation Fusion, 14(1):28–44, 2013.
62 Iuliia Kotseruba, John K. Tsotsos

277. David Kieras. Modeling Visual Search of Displays of Many Objects: The Role of Differential Acuity and Fixation Memory.
Proceedings of the 10th international conference on cognitive modeling, 2010.
278. David Kieras. The Control of Cognition. In W. Gray, editor, Integrated Models of Cognitive Systems. Oxford University
Press, 2012.
279. David E. Kieras. EPIC Architecture Principles of Operation, 2004.
280. David E. Kieras and Anthony J. Hornof. Towards accurate and practical predictive models of active-vision-based visual
search. In Proceedings of the Conference on Human Factors in Computing Systems, pages 3875–3884, 2014.
281. David E. Kieras and David E. Meyer. The Role of Cognitive Task Analysis in the Application of Predictive Models of
Human Performance. EPIC Report No. 11 (TR-98/ONR-EPIC-11), 1998.
282. David E. Kieras, Gregory H. Wakefield, Eric R. Thompson, Nandini Iyer, and Brian D. Simpson. Modeling Two-Channel
Speech Processing With the EPIC Cognitive Architecture. Topics in Cognitive Science, 8(1):291–304, 2016.
283. Ozkan Kilic. Intelligent Reasoning on Natural Language Data: A Non-Axiomatic Reasoning System Approach. PhD
Thesis, 2015.
284. David Kinny, Michael Georgeff, and James Hendler. Experiments in Optimal Sensing for Situated Agents. In Proceedings
of the Second Pacific Rim International Conference on Artificial Intelligence, 1992.
285. James R. Kirk and John E. Laird. Interactive Task Learning for Simple Games. Advances in Cognitive Systems, 3:13–30,
2014.
286. James R. Kirk and John E. Laird. Learning General and Efficient Representations of Novel Games Through Interactive
Instruction. Advances in Cognitive Systems, 4, 2016.
287. Kiril Kiryazov, Georgi Petkov, Maurice Grinberg, Boicho Kokinov, and Christian Balkenius. The Interplay of Analogy-
Making with Active Vision and Motor Control in Anticipatory Robots. Workshop on Anticipatory Behavior in Adaptive
Learning Systems, pages 233–253, 2007.
288. Boicho Kokinov, Vassil Nikolov, and Alexander Petrov. Dynamics of emergent computation in DUAL. Artificial intelli-
gence: Methodology, Systems, Applications, 1996.
289. Boicho Nikolov Kokinov. Associative Memory-Based Reasoning: Some Experimental Results. In Proceedings of the twelfth
annual conference of the Cognitive Science Society, 1990.
290. Boicho Nikolov Kokinov. The DUAL Cognitive Architecture: A Hybrid Multi- Agent Approach. In Proceedings of the
11th European conference on artificial intelligence (ECAI), 1994.
291. Robert Koons. Defeasible Reasoning, 2017.
292. Ioannis Kostavelis, Lazaros Nalpantidis, and Antonios Gasteratos. Object recognition using saliency maps and HTM
learning. In Proceedings of the IEEE International Conference on Imaging Systems and Techniques, pages 528–532, 2012.
293. Iuliia Kotseruba. Visual Attention in Dynamic Environments and Its Application To Playing Online Games. MSc Thesis,
2016.
294. Jonathan Kottlors, Daniel Brand, and Marco Ragni. Modeling Behavior of Attention-Deficit-Disorder Patients in a N-Back
Task. In Proceedings of 11th International Conference on Cognitive Modeling (ICCM 2012), pages 297–302, 2012.
295. Jeffrey L. Krichmar. Design principles for biologically inspired cognitive robotics. Biologically Inspired Cognitive Archi-
tectures, 1:73–81, 2012.
296. Jeffrey L. Krichmar and Gerald M. Edelman. Brain-based devices for the study of nervous systems and the development
of intelligent machines. Artificial Life, 11(1-2):63–77, 2005.
297. Jeffrey L. Krichmar, Douglas A. Nitz, Joseph A. Gally, and Gerald M. Edelman. Characterizing functional hippocampal
pathways in a brain-based device as it solves a spatial memory task. Proceedings of the National Academy of Sciences of
the United States of America, 102(6):2111–2116, 2005.
298. Jeffrey L. Krichmar and James A. Snook. A neural approach to adaptive behavior and multi-sensor action selection in a
mobile device. In Proceedings of the IEEE International Conference on Robotics and Automation, 2002.
299. Daniel R Kuokka. Integrating Planning, Execution, and Learning. In Proceedings of the NASA Conference on Space
Telerobotics, pages 377–386, 1989.
300. Daniel R. Kuokka. MAX: A Meta-Reasoning Architecture for ”X”. SIGART Bulletin, 2(4):93–97, 1991.
301. Unmesh Kurup, Perrin G. Bignoli, J. R. Scally, and Nicholas L. Cassimatis. An architectural framework for complex
cognition. Cognitive Systems Research, 12(3-4):281–292, 2011.
302. J. E. Laird, K. J. Coulter, R. M. Jones, P. G. Kenny, Frank Koss, and P. E. Nielsen. Integrating intelligent computer gener-
ated forces in distributed simulations: TacAir-Soar in STOW-97. In Proceedings of the Spring Simulation Interoperability
Workshop, 1998.
303. J. E. Laird, C. Lebiere, and P. S. Rosenbloom. A Standard Model for the Mind: Toward a Common Computational
Framework across Artificial Intelligence, Cognitive Science, Neuroscience, and Robotics. AI Magazine, In press, 2017.
304. J. E. Laird and S. Mohan. A case study of knowledge integration across multiple memories in Soar. Biologically Inspired
Cognitive Architectures, 8:93–99, 2014.
305. J. E. Laird, P. S. Rosenbloom, and A. Newell. Towards chunking as a general learning mechanism. In AAAI Proceedings,
pages 188–192, 1984.
A Review of 40 Years in Cognitive Architecture Research Core Cognitive Abilities and Practical Applications 63

306. John E Laird. The Soar Cognitive Architecture. MIT Press, 2012.
307. John E. Laird. The Soar Cognitive Architecture. AISB Quarterly, 171(134):224–235, 2012.
308. John E. Laird and Nate Derbinsky. A Year of Episodic Memory. In Proceedings of the Workshop on Grand Challenges
for Reasoning from Experiences, IJCAI, pages 7–10, 2009.
309. John E. Laird, Keegan R. Kinkade, Shiwali Mohan, and Joseph Z. Xu. Cognitive Robotics Using the Soar Cognitive
Architecture. In Proceedings of the 6th International Conference on Cognitive Modelling, pages 226–330, 2004.
310. John E. Laird, Eric S. Yager, Michael Hucka, and Christopher M. Tuck. Robo-Soar: An integration of external interaction,
planning, and learning using Soar. Robotics and Autonomous Systems, 8(1-2):113–129, 1991.
311. K Landauer. How Much Do People Remember ! Some Estimates of the Quantity of Learned Information in Long-term
Memory. Cognitive Science, 493:477–493, 1986.
312. Peter C R Lane, Fernand Gobet, and Richard Ll Smith. Attention mechanisms in the CHREST cognitive architecture.
Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in
Bioinformatics), 5395 LNAI:183–196, 2009.
313. Peter C. R. Lane, Anthony Sykes, and Fernand Gobet. Combining Low-Level Perception with Expectations in CHREST.
In Proceedings of the European Cognitive Science Conference, pages 205–210, 2003.
314. Pat Langley and John A. Allen. A Unified Framework for Planning and Learning. In S. Minton, editor, Machine Learning
methods for Planning. Morgan Kaufmann, 1993.
315. Pat Langley, Dongkyu Choi, and Seth Rogers. Interleaving Learning, Problem Solving, and Execution in the ICARUS
Architecture. Technical Report, Computational Learning Laboratory, 2005.
316. Pat Langley, Kirstin Cummings, and Daniel Shapiro. Hierarchical skills and cognitive architectures. In Proceedings of the
26th Annual Conference of the Cognitive Science Society, pages 779–784, 2004.
317. Pat Langley, John E. Laird, and Seth Rogers. Cognitive architectures: Research issues and challenges. Cognitive Systems
Research, 10(2):141–160, 2009.
318. Pat Langley and Seth Rogers. An Extended Theory of Human Problem Solving. In Proceedings of the 27th Annual
Meeting of Cognitive Science Society, pages 166–186, 2008.
319. A. Lavin, S. Ahmad, and J. Hawkins. Sparse Distributed Representations, 2016.
320. Alexander Lavin and Subutai Ahmad. Evaluating Real-time Anomaly Detection Algorithms – the Numenta Anomaly
Benchmark. In Proceedings of the14th International Conference on Machine Learning and Applications (ICMLA), 2015.
321. Christian Lebiere, Peter Pirolli, Robert Thomson, Jaehyon Paik, Matthew Rutledge-Taylor, James Staszewski, and John R.
Anderson. A functional model of sensemaking in a neurocognitive architecture. Computational Intelligence and Neuro-
science, 2013, 2013.
322. Christian (Carnegie Mellon University) Lebiere, Eric (Carnegie Mellon University) Biefeld, Rick (Micro Analysis & De-
sign) Archer, Sue (Micro Analysis & Design) Archer, Laurel (Army Research Laboratory) Allender, and Troy D. (Army
Research Laboratory) Kelley. IMPRINT/ACT-R: Integration of a task network modeling architecture with a cognitive
architecture and its application to human error modeling. Proceedings of the 2002 Advanced Simulation Technologies
Conference, San Diego, CA, Simulation Series, 34:13–19, 2002.
323. Shane Legg and Marcus Hutter. A Collection of Definitions of Intelligence. Frontiers in Artificial Intelligence and
applications, 157, 2007.
324. Jorgen Leitner, Simon Harding, Mikhail Frank, Alexander Forster, and Jorgen Schmidhuber. An integrated, modular
framework for computer vision and cognitive robotics research (icVision). Advances in Intelligent Systems and Computing,
pages 205–210, 2013.
325. Itamar Lerner, Shlomo Bentin, and Oren Shriki. Spreading Activation in an Attractor Network with Latching Dynamics:
Automatic Semantic Priming Revisited. Cognitive Science, 36(8):1339–1382, 2012.
326. R. L. Lewis. Recent developments in the NL-Soar garden path theory. Technical Report CMU-CS-93-141, 1992.
327. Ruiting Lian, Ben Goertzel, Rui Liu, Michael Ross, Murilo Queiroz, and Linas Vepstas. Sentence generation for artificial
brains: A glocal similarity-matching approach. Neurocomputing, 74(1-3):95–103, 2010.
328. Antonio Lieto. Representational Limits in Cognitive Architectures. In Proc. EUCognition, volume 1855, 2016.
329. Antonio Lieto, Christian Lebiere, and Alessandro Oltramari. The Knowledge Level in Cognitive Architectures: Current
Limitations and Possible Developments. Cogn. Syst. Res., In Press, 2017.
330. Peter Lindes and John E Laird. Toward Integrating Cognitive Linguistics and Cognitive Language Processing. In Pro-
ceedings of International Conference on Cognitive Modeling, 2016.
331. Pierre Lison and Geert-Jan Kruijff. Salience-driven Contextual Priming of Speech Recognition for Human-Robot Interac-
tion. In Language, pages 636–640, 2008.
332. Joan Marc Llargues Asensio, Juan Peralta, Raul Arrabales, Manuel Gonzalez Bedia, Paulo Cortez, and Antonio Lopez
Pe??a. Artificial Intelligence approaches for the generation and assessment of believable human-like behaviour in virtual
characters. Expert Systems with Applications, 41(16):7281–7290, 2014.
333. Martyn Lloyd-Kelly, Fernand R. Gobet, and Peter C. R. Lane. Piece of Mind: Long-Term Memory Structure in ACT-R
and CHREST. In Proceedings of the 37th Annual Meeting of the Cognitive Science Society, 2015.
64 Iuliia Kotseruba, John K. Tsotsos

334. Martyn Lloyd-Kelly, Peter C. R. Lane, and Fernand Gobet. The Effects of Bounding Rationality on the Performance and
Learning of CHREST Agents in Tileworld. In Max Bramer and Miltos Petridis, editors, Research and Development in
Intelligent Systems XXXI. 2014.
335. Dorian Gálvez López, Kristoffer Sjö, Chandana Paul, and Patric Jensfelt. Hybrid laser and vision based object search and
localization. Proceedings of the IEEE International Conference on Robotics and Automation (ICRA), 2008.
336. Alan M. Lytle and Kamel S. Saidi. NIST research in autonomous construction. Autonomous Robots, 22(3):211–221, 2007.
337. Natalia Lyubova, David Filliat, and Serena Ivaldi. Improving object learning through manipulation and robot self-
identification. In Proceeding of the IEEE International Conference on Robotics and Biomimetics (ROBIO), 2013.
338. Tamas Madl and Stan Franklin. A LIDA-based model of the attentional blink. In Proceedings of International Conference
on Cognitive Modeling (ICCM), pages 283–288, 2012.
339. Tamas Madl and Stan Franklin. Constrained Incrementalist Moral Decision Making for a Biologically Inspired Cognitive
Architecture. In Robert Trappl, editor, A Construction Manual for Robots’ Ethical Systems. 2015.
340. Tamas Madl, Stan Franklin, Ke Chen, Daniela Montaldi, and Robert Trappl. Towards real-world capable spatial memory
in the LIDA cognitive architecture. Biologically Inspired Cognitive Architectures, 16, 2015.
341. Giovanni Maffei, Diogo Santos-Pata, Encarni Marcos, Marti Sánchez-Fibla, and Paul F M J Verschure. An embodied
biologically constrained model of foraging: From classical and operant conditioning to adaptive real-world behavior in
DAC-X. Neural Networks, 72:88–108, 2015.
342. Xiaochun Mai, Xinzheng Zhang, Yichen Jin, Yi Yang, and Jianfen Zhang. Simple Perception-Action Strategy Based
on Hierarchical Temporal Memory. In Proceeding of the IEEE International Conference on Robotics and Biomimetics
(ROBIO), pages 1759–1764, 2013.
343. L. J. Manso, L. V. Calderita, P. Bustos, J. Garcia, M. Martinez, F. Fernandez, A. Romero-Garces, and A. Bandera. A
General-Purpose Architecture to Control Mobile Robots. In XV Workshop of physical agents: Book of proceedings (WAF
2014), 2014.
344. Jonatas Manzolli and Paul F.M.J. Verschure. Roboser: A Real-World Composition System. pages 0–9, 2005.
345. Robert P. Marinier and John E. Laird. Toward a Comprehensive Computational Model of Emotions and Feelings. In
Proceedings of Sixth International Conference on Cognitive Modeling: ICCM, pages 172–177, 2004.
346. Robert P. Marinier, John E. Laird, and Richard L. Lewis. A computational unification of cognitive behavior and emotion.
Cognitive Systems Research, 10(1):48–69, 2009.
347. D. Marr. Vision: A Computational Investigation Into the Human Representation and Processing of Visual Information.
MIT Press, 2010.
348. James B. Marshall. Metacat: A Program That Judges Creative Analogies in a Microworld. In Proceedings to the Second
Workshop on Creative Systems, 2002.
349. James B. Marshall. Metacat: A Self-Watching Cognitive Architecture for Analogy-Making. In Proceedings of the 24th
Annual Conference of the Cognitive Science Society, 2002.
350. James B. Marshall. A self-watching model of analogy-making and perception. Journal of Experimental & Theoretical
Artificial Intelligence, 18(3):267–307, 2006.
351. Siegfried Martens, Gail A. Carpenter, and Paolo Gaudiano. Neural sensor fusion for spatial visualization on a mobile
robot, 1998.
352. D. Martin, M. Rincon, M. C. Garcia-Alegre, and D. Guinea. ARDIS: Knowledge-based dynamic architecture for real-time
surface visual inspection. Lecture Notes in Computer Science, 5601:395–404, 2009.
353. D. Martin, M. Rincon, M. C. Garcia-Alegre, and D. Guinea. ARDIS: Knowledge-based architecture for visual system
configuration in dynamic surface inspection. Expert Systems, 28(4):353–374, 2011.
354. Jesus Martinez-Gomez, Rebeca Marfil, Luis V. Calderita, Juan Pedro Bandera, Luis J. Manso, Antonio Bandera, Adrian
Romero-Garces, and Pablo Bustos. Toward Social Cognition in Robotics: Extracting and Internalizing Meaning from
Perception. In Workshop of Physical Agents, 2014.
355. Joao Martins and Vilela Mendes. Neural networks and logical reasoning systems. A translation table. International
Journal of Neural Systems, 11(2):179–186, 2001.
356. Zenon Mathews, Sergi Bermúdez I Badia, and Paul F M J Verschure. PASAR: An integrated model of prediction,
anticipation, sensation, attention and response for artificial sensorimotor systems. Information Sciences, 186(1):1–19,
2012.
357. Zenon Mathews, Miguel Lechon, J. M. Blanco Calvo, Anant Dhir Armin Duff, Sergi Bermudez I. Badia, and Paul F. M. J.
Verschure. Insect-like mapless navigation based on head direction cells and contextual learning using chemo-visual sensors.
2009 IEEE/RSJ International Conference on Intelligent Robots and Systems, IROS 2009, pages 2243–2250, 2009.
358. Larry Matthies. Stereo vision for planetary rovers: Stochastic modeling to near real-time implementation. International
Journal of Computer Vision, 8(1):71–91, 1992.
359. James B. Maxwell. Generative Music, Cognitive Modelling, and Computer-Assisted Composition in MusiCog and ManuS-
core. PhD Thesis, 2014.
A Review of 40 Years in Cognitive Architecture Research Core Cognitive Abilities and Practical Applications 65

360. James B. Maxwell, Arne Eigenfeldt, Philippe Pasquier, and Nicolas Gonzalez Thomas. Musicog: a Cognitive Architecture
for Music Learning and Generation. In Proceedings of the 9th Sound and Music Computing Conference, pages 521–528,
2012.
361. Robert R McCrae and Oliver P John. An introduction to the five-factor model and its applications. Journal of personality,
60(2):175–215, 1992.
362. Harry McGurk and John MacDonald. Hearing lips and seeing voices. Nature, 264(5588):746–748, 1976.
363. Wim J. C. Melis, Shuhei Chizuwa, and Michitaka Kameyama. Evaluation of hierarchical temporal memory for a real
world application. Proceedings of the 4th International Conference on Innovative Computing, Information and Control
(ICICIC), pages 144–147, 2009.
364. David H. Menager and Dongkyu Choi. A Robust Implementation of Episodic Memory for a Cognitive Architecture.
Proceedings of Annual Meeting of the Cognitive Science Society, pages 620–625, 2016.
365. Giorgio Metta, Lorenzo Natale, Francesco Nori, Giulio Sandini, David Vernon, Luciano Fadiga, Claes von Hofsten, Kerstin
Rosander, Manuel Lopes, Jose Santos-Victor, Alexandre Bernardino, and Luis Montesano. The iCub humanoid robot: An
open-systems platform for research in cognitive development. Neural Networks, 23(8-9):1125–1134, 2010.
366. Tomas Mikolov, Armand Joulin, and Marco Baroni. A Roadmap towards Machine Intelligence. arXiv : 1511.08130v1,
2015.
367. David P. Miller and Marc G. Slack. Global symbolic maps from local navigation. In Proceedings of the ninth National
conference on Artificial intelligence AAAI, pages 750–755, 1991.
368. Aaron Mininger and John Laird. Interactively Learning Strategies for Handling References to Unseen or Unknown Objects.
Advances in Cognitive Systems, 5, 2016.
369. Marvin Minsky. The Society of Mind. Simon & Shuster, Inc., New York, NY, 1986.
370. Steven Minton, Jaime Carbonell, Craig A. Knoblock, Daniel R. Kuokka, Oren Etzioni, and Yolanda Gil. Explanation-Based
Learning: A Problem Solving Perspective. Artificial Intelligence, 40(1–3):63–118, 1989.
371. D. K. Mitchell. Workload Analysis of the Crew of the Abrams V2 SEP: Phase I Baseline IMPRINT Model. Technical
Report ARL-TR-5028, 2009.
372. D. K. Mitchell, Brooke Abounader, and Shanell Henry. A Procedure for Collecting Mental Workload Data During an
Experiment That Is Comparable to IMPRINT Workload Data. Tehcnical Report ARL-TR-5020, 2009.
373. Melanie Mitchell and Douglas R. Hofstadter. The emergence of understanding in a computer model of concepts and
analogy-making. Physica D: Nonlinear Phenomena, 42(1-3):322–334, 1990.
374. Tom Mitchell, John Allen, Prasad Chalasani, John Cheng, Oren Etzioni, Marc Ringuette, and Jeffrey C. Schlimmer. Theo:
A framework for self-improving systems. In K. VanLehn, editor, Architectures for intelligence, pages 323–356. Erbaum,
1989.
375. Tom M. Mitchell. Becoming Increasingly Reactive. In Proceedings of the Eighth National Conference on Artificial
Intelligence, pages 1051–1058, 1990.
376. Volodymyr Mnih, Koray Kavukcuoglu, David Silver, Andrei A. Rusu, Joel Veness, Marc G. Bellemare, Alex Graves, Martin
Riedmiller, Andreas K. Fidjeland, Georg Ostrovski, Stig Petersen, Charles Beattie, Amir Sadik, Ioannis Antonoglou,
Helen King, Dharshan Kumaran, Daan Wierstra, Shane Legg, and Demis Hassabis. Human-level control through deep
reinforcement learning. Nature, 518(7540):529–533, 2015.
377. Shiwali Mohan, Aaron H. Mininger, James R. Kirk, and John E. Laird. Acquiring Grounded Representations of Words
with Situated Interactive Instruction. Advances in Cognitive Systems, 2:113–130, 2012.
378. Jungaa Moon and John R. Anderson. Timing in multitasking: Memory contamination and time pressure bias. Cognitive
Psychology, 67:26–54, 2013.
379. Antonio Miguel Mora, Francisco Aisa, Pablo Garcı́a-Sánchez, Pedro Ángel Castillo, and Juan Julián Merelo. Modelling a
Human-Like Bot in a First Person Shooter Game. International Journal of Creative Interfaces and Computer Graphics,
6(1):21–37, 2015.
380. Shane T. Mueller and Brandon S. Minnery. Adapting the Turing Test for Embodied Neurocognitive Evaluation of
Biologically-Inspired cognitive agents. In Proc. AAAI Fall Symp. Biol. Inspired Cogn. Archit., 2008.
381. J. William Murdock and Ashok K. Goel. Meta-case-based reasoning: self-improvement through self-understanding. Journal
of Experimental & Theoretical Artificial Intelligence, 20(1):1–36, 2008.
382. William Murdock and Ashok Goel. Meta-case-Based Reasoning : Using Functional Models to Adapt Case-Based Agents.
In Proceedings of the 4th. International Conference on Case-Based Reasoning, 2001.
383. K. N. Murphy, R. J. Norcross, and F. M. Proctor. CAD directed robotic deburring. In Proceedings of the second
international symposium on robotics and manufacturing research, education, and applications, 1988.
384. Isabel Briggs Myers, Mary H McCaulley, Naomi L Quenk, and Allen L Hammer. MBTI manual: A guide to the development
and use of the Myers-Briggs Type Indicator, volume 3. Consulting Psychologists Press Palo Alto, CA, 1998.
385. Karen L. Myers, David L. Martin, and David N. Morley. Taskable Reactive Agent Communities. Final Technical Report
AFRL-IF-RS-TR-2002-208, 2002.
386. Allen Newell. Physical symbol systems. Cognitive Science, 4(2), 1980.
66 Iuliia Kotseruba, John K. Tsotsos

387. Allen Newell. Précis of Unified theories of cognition. Behavioral and Brain Sciences, 15:425–492, 1992.
388. G. W. Ng, Xuhong Xiao, R. Z. Chan, and Y. S. Tan. Scene Understanding using DSO Cognitive Architecture. In
Proceedings of the 15th International Conference on Information Fusion (FUSION), pages 2277–2284, 2012.
389. Sao Mai Nguyen, Serena Ivaldi, Natalia Lyubova, Alain Droniou, Damien Gerardeaux-Viret, David Filliat, Vincent Padois,
Olivier Sigaud, and Pierre Yves Oudeyer. Learning to recognize objects through curiosity-driven manipulation with the
iCub humanoid robot. In Proceedings of the 3rd Joint International Conference on Development and Learning and
Epigenetic Robotics, 2013.
390. Yael Niv. Reinforcement learning in the brain. Journal of Mathematical Psychology, 53(3), 2009.
391. Rony Novianto. Flexible Attention-based Cognitive Architecture for Robots. PhD Thesis, 2014.
392. Rony Novianto, Benjamin Johnston, and Mary Anne Williams. Attention in the ASMO cognitive architecture. Frontiers
in Artificial Intelligence and Applications, 221:98–105, 2010.
393. Rony Novianto, Benjamin Johnston, and Mary Anne Williams. Habituation and sensitisation learning in ASMO cognitive
architecture. Lecture Notes in Computer Science, 8239 LNAI:249–259, 2013.
394. P. Nunez, Luis J. Manso, Pablo Bustos, Paulo Drews-Jr, and Douglas G. Macharet. Towards a new Semantic Social
Navigation Paradigm for Autonomous Robots using CORTEX. In IEEE International Symposium on Robot and Human
Interactive Communication (RO-MAN 2016) - BAILAR2016 Workshop, 2016.
395. Andrew M. Nuxoll and John E. Laird. Extending Cognitive Architecture with Episodic Memory. In Proceedings of the
National Conference on Artificial Intelligence., 2007.
396. E. Nyamsuren and Niels A. Taatgen. Human Reasoning Module. Biologically Inspired Cognitive Architectures, 8, 2014.
397. Enkhbold Nyamsuren and Niels A. Taatgen. Pre-attentive and attentive vision module. Cognitive Systems Research,
pages 211–216, 2013.
398. Gary H. Ogasawara. A Distributed, Decision-Theoretic Control System for a Mobile Robot. SIGART Bulletin, 2(4):140–
145, 1991.
399. Gary H. Ogasawara and Stuart J. Russell. Planning Using Multiple Execution Architectures. In Proceedings of the
International Joint Conference on Artificial Intelligence, 1993.
400. R. C. O’Reilly. Modeling integration and dissociation in brain and cognitive development. In Processes of change in brain
and cognitive development: Attention and performance XXI, pages 375–401. 2006.
401. Randall C. O’Reilly and Michael J. Frank. Making working memory work: a computational model of learning in the
prefrontal cortex and basal ganglia. Neural computation, 18(2):283–328, 2006.
402. Randall C. O’Reilly, Thomas E. Hazy, and Seth A. Herd. The Leabra cognitive architecture: how to play 20 principles
with nature and win! In The Oxford Handbook of Cognitive Science, pages 1–31. 2012.
403. Randall C. O’Reilly, Thomas E. Hazy, Jessica Mollick, Prescott Mackie, and Seth Herd. Goal-Driven Cognition in the
Brain: A Computational Framework. arXiv preprint arXiv:1404.7591, 2014.
404. Randall C. O’Reilly and James H. Hoeffner. Competition, Priming, and the Past Tense U-Shaped Developmental Curve.
2000.
405. Randall C. O’Reilly, Dean Wyatte, Seth Herd, Brian Mingus, and David J. Jilk. Recurrent processing during object
recognition. Frontiers in Psychology, 4, 2013.
406. Pinar Ozturk. Levels and Types of Action Selection: The Action Selection Soup. Adaptive Behavior, 2009.
407. Elena Pacchierotti, Henrik I. Christensen, and Patric Jensfelt. Embodied social interaction for service robots in hallway
environments. In Proceedings of the 5th International Conference on Field and Service Robotics, 2005.
408. Matt Paisner, Michael T. Cox, Michael Maynord, and Don Perlis. Goal-Driven Autonomy for Cognitive Systems. In
Proceedings of the 36th Annual Conference of the Cognitive Science Society, pages 2085–2090, 2013.
409. Nele Pape and Leon Urbas. A Model of Time-Estimation Considering Working Memory Demands. Proceedings of the
30th annual conference of the Cognitive Science Society, pages 1543–1548, 2008.
410. Ugo Pattacini, Francesco Nori, Lorenzo Natale, Giorgio Metta, and Giulio Sandini. An experimental evaluation of a novel
minimum-jerk Cartesian controller for humanoid robots. In Proceedings of the International Conference on Intelligent
Robots and Systems (IROS), pages 1668–1674, 2010.
411. Andreas Perner and Heimo Zeilinger. Action primitives for bionics inspired action planning system: Abstraction layers
for action planning based on psychoanalytical concepts. In IEEE International Conference on Industrial Informatics
(INDIN), pages 63–68, 2011.
412. R. Alan II Peters, Kimberly A. Hambuchen, Kazuhiko Kawamura, and D. Mitchell Wilkes. The Sensory Ego-Sphere as
a Short-Term Memory for Humanoids. In Proceedings of the IEEE-RAS International Conference on Humanoid Robots,
2001.
413. Richard Alan Peters, Kazuhiko Kawamura, D. Mitchell Wilkes, Kimberly A. Hambuchen, Tamara E. Rogers, and W. An-
thony Alford. ISAC Humanoid: An Architecture for Learning and Emotion. In Proceedings of the IEEE-RAS International
Conference on Humanoid Robots, number 1, page 459, 2001.
414. Georgi Petkov, Tchavdar Naydenov, Maurice Grinberg, and Boicho Kokinov. Building Robots with Analogy-Based An-
ticipation. Annual Conference on Artificial Intelligence, 2006.
A Review of 40 Years in Cognitive Architecture Research Core Cognitive Abilities and Practical Applications 67

415. Giovanni Pezzulo. DiPRA: a layered agent architecture which integrates practical reasoning and sensorimotor schemas.
Connection Science, 21(4):297–326, 2009.
416. Giovanni Pezzulo and Gianguglielmo Calvi. Dynamic Computation and Context Effects in the Hybrid Architecture
AKIRA. International and Interdisciplinary Conference on Modeling and Using Context., 2005.
417. Giovanni Pezzulo, Gianguglielmo Calvi, and Cristiano Castelfranchi. DiPRA: Distributed Practical Reasoning Architec-
ture. In Proceedings of International Joint Conference on Artificial Intelligence (IJCAI), pages 1458–1463, 2007.
418. Andrew B. Philips and John L. Bresina. NASA TileWorld. NASA Technical Report TR-FIA-91-04, 1991.
419. Paolo Pirjanian. Behavior Coordination Mechanisms. Technical Report IRIS-99-375, 1999.
420. John L. Pollock. Oscar - A General-Purpose Defeasible Reasoner. AAAI Technical Report FS-93-01, 1993.
421. John L. Pollock. Planning in OSCAR. Minds and Machines, 2:113–144, 1993.
422. John L. Pollock. OSCAR: An Agent Architecture Based on Defeasible Reasoning. In AAAI Spring Symposium: Emotion,
Personality, and Social Behavior, 2008.
423. John L. Pollock and Devin Hosea. OSCAR-MDA: An Artificially Intelligent Advisor for Emergency Room Medicine. 1995.
424. Michael Posner, Mary Jo NIssen, and Raymond M. Klein. Visual dominance: An information-processing account of its
origins and significance. Psychological Review, 83(2):157–171, 1976.
425. Stefan Profanter. Cognitive architectures. In Hauptseminar Human Robot Interaction, 2012.
426. David V. Pynadath, Paul S. Rosenbloom, and Stacy C. Marsella. Reinforcement Learning for Adaptive Theory of Mind
in the Sigma Cognitive Architecture. In Proceedings of the Conference on Artificial General Intelligence, 2014.
427. David V. Pynadath, Paul S. Rosenbloom, Stacy C. Marsella, and Lingshan Li. Modeling two-player games in the sigma
graphical cognitive architecture. In International Conference on Artificial General Intelligence, 2013.
428. Catharine H. Rankin, Thomas Abrams, Robert J. Barry, Seema Bhatnagar, David Clayton, John Colombo, Gianluca
Coppola, Mark A. Geyer, David L. Glanzman, Stephen Marsland, Frances Mcsweeney, Donald A. Wilson, Chun-Fang Wu,
and Richard F. Thompson. Habituation Revisited: An Updated and Revised Description of the Behavioral Characteristics
of Habituation. Neurobiology of Learning and Memory, 92(2):135–138, 2009.
429. Anand S. Rao and Michael P. George. Intelligent Real-Time Network Management. In Proceedings of the Tenth Interna-
tional Conference on AI, Expert Systems and Natural Language, 1991.
430. Daniel Rasmussen and Chris Eliasmith. Modeling Brain Function Current Developments and Future Prospects. JAMA
Neurology, 70(10):1325–1329, 2013.
431. Rainer Reisenzein, Eva Hudlicka, Mehdi Dastani, Jonathan Gratch, Koen Hindriks, Emiliano Lorini, and John Jules Ch.
Meyer. Computational modeling of emotion: Toward improving the inter- and intradisciplinary exchange. IEEE Trans-
actions on Affective Computing, 4(3):246–266, 2013.
432. Maximilian Riesenhuber. Object recognition in cortex: Neural mechanisms, and possible roles for attention. Neurobiology
of Attention, 2005.
433. Frank E. Ritter. Two Cognitive Modeling Frontiers. Emotions and Usability. Information and Media Technologies,
4(1):76–84, 2009.
434. Frank E. Ritter, Jennifer L. Bittner, Sue E. Kase, Rick Evertsz, Matteo Pedrotti, and Paolo Busetta. CoJACK: A
high-level cognitive architecture with demonstrations of moderators, variability, and implications for situation awareness.
Biologically Inspired Cognitive Architectures, 1:2–13, 2012.
435. Brandon Rohrer. A developmental agent for learning features, environment models, and general robotics tasks.
ICDL/Eprirob, 2011.
436. Brandon Rohrer. An implemented architecture for feature creation and general reinforcement learning. In Fourth Inter-
national Conference on Artificial General Intelligence, Workshop on Self-Programming in AGI Systems, 2011.
437. Brandon Rohrer. Biologically inspired feature creation for multi-sensory perception. Biologically Inspired Cognitive
Architectures, pages 305–313, 2011.
438. Brandon Rohrer. BECCA: Reintegrating AI for natural world interaction. In AAAI Spring Symposium: Designing
Intelligent Robots. AAAI Technical Report SS-12-02, 2012.
439. Brandon Rohrer. BECCA version 0.4.5. User’s Guide, 2013.
440. Brandon Rohrer, Michael Bernard, Dan J. Morrow, Fred Rothganger, and Patrick Xavier. Model-free learning and control
in a mobile robot. In Proceedings of the 5th International Conference on Natural Computation, ICNC 2009, pages 566–572,
2009.
441. Adrián Romero-Garcés, Luis Vicente Calderita, Jesús Martı́nez-Gómez, Juan Pedro Bandera, Rebeca Marfil, Luis J.
Manso, Antonio Bandera, and Pablo Bustos. Testing a fully autonomous robotic salesman in real scenarios. In IEEE
International Conference on Autonomous Robots Systems and Competitions, 2015.
442. Adrián Romero-Garcés, Luis Vicente Calderita, Jesus Martinez-Gomez, Juan Pedro Bandera, Rebeca Marfil, Luis J.
Manso, Pablo Bustos, and Antonio Bandera. The cognitive architecture of a robotic salesman. Conference of the Spanish
Association for Artificial Intelligence, 15(6), 2015.
443. Paul S. Rosenbloom, Abram Demski, and Volkan Ustun. Efficient message computation in Sigma’s graphical architecture.
Biologically Inspired Cognitive Architectures, 11:1–9, 2015.
68 Iuliia Kotseruba, John K. Tsotsos

444. Paul S. Rosenbloom, Jonathan Gratch, and Volkan Ustun. Towards Emotion in Sigma: From Appraisal to Attention. In
International Conference on Artificial General Intelligence, 2015.
445. Paul S. Rosenbloom, John E. Laird, Allen Newell, and Robert McCarl. A preliminary analysis of the Soar architecture as
a basis for general intelligence. Artificial Intelligence, 47(1-3):289–325, 1991.
446. Casey Rosenthal and Clare Bates Congdon. Personality profiles for generating believable bot behaviors. In Proceedings
of the IEEE Conference on Computational Intelligence and Games, pages 124–131, 2012.
447. Daniel Rousseau and Barbara Hayes-Roth. Personality in Synthetic Agents. Report No. KSL 96-21, 1996.
448. Daniel Rousseau and Barbara Hayes-roth. Interacting with Personality-Rich Characters. Report No. KSL 97-06, 1997.
449. Jonas Ruesch, Manuel Lopes, Alexandre Bernardino, Jonas Hornstein, Jose Santos-Victor, and Rolf Pfeifer. Multimodal
saliency-based bottom-up attention a framework for the humanoid robot iCub. In Proceedings of the IEEE International
Conference on Robotics and Automation, pages 962–967, 2008.
450. D. Ruiz and A. Newell. Tower-noticing triggers strategy-change in the Tower of Hanoi: A Soar model. Technical Report
AIP-66, pages 522–529, 1989.
451. Stuart J. Russel and Eric Wefald. Decision-Theoretic Control of Reasoning: General Theory and an Application to
Game-Playing. Technical Report UCB/CSD 88/435, 1988.
452. Stuart Russell and Peter Norvig. Artificial Intelligence: A Modern Approach. Prentice Hall, 1995.
453. Stuart Russell and Eric Wefald. On Optimal Game-Tree Search using Rational Meta-Reasoning. In Proceedings of the
International Joint Conference on Artificial Intelligence, 1989.
454. R. Salgado, F. Bellas, P. Caamano, B. Santos-Diez, and R. J. Duro. A procedural Long Term Memory for cognitive
robotics. In Proceedings of the IEEE Conference on Evolving and Adaptive Intelligent Systems, pages 57–62, 2012.
455. Dario D. Salvucci. A Model of Eye Movements and Visual Attention. Proceedings of the Third International Conference
on Cognitive Modeling, pages 252–259, 2000.
456. Alexei V. Samsonovich. Toward a unified catalog of implemented cognitive architectures. In Proceeding of the Conference
on Biologically Inspired Cognitive Architectures, pages 195–244, 2010.
457. Alexei V Samsonovich, Giorgio a Ascoli, Kenneth a De Jong, and Mark a Coletti. Integrated Hybrid Cognitive Architecture
for a Virtual Roboscout. In M. Beetz, K. Rajan, M. Thielscher, and R.B. Rusu, editors, Cognitive robotics: Papers from
the AAAI workshop, AAAI technical reports, volume 6, pages 129–134. AAAI Press, 2006.
458. Alexei V. Samsonovich, Kenneth A. De Jong, Anastasia Kitsantas, Erin E. Peters, Nada Dabbagh, and M. Layne
Kalbfleisch. Cognitive constructor: An intelligent tutoring system based on a biologically inspired cognitive architec-
ture (BICA). Frontiers in Artificial Intelligence and Applications, 171:311–325, 2008.
459. Giulio Sandini, Giorgio Metta, and David Vernon. The iCub Cognitive Humanoid Robot: An Open-System Research
Platform for Enactive Cognition. Lecture Notes in Computer Science, pages 358–369, 2007.
460. Scott Sanner, John R. Anderson, Christian Lebiere, and Marsha C. Lovett. Achieving Efficient and Cognitively Plausible
Learning in Backgammon. In Proceedings of the Seventeenth International Conference on Machine Learning (ICML-
2000), 2000.
461. Scott P Sanner. A Quick Introduction to 4CAPS Programming, 1999.
462. John F. Santore and Stuart C. Shapiro. Crystal Cassie: Use of a 3-D Gaming Environment for a Cognitive Agent. Papers
of the IJCAI 2003 Workshop on Cognitive Modeling of Agents and Multi-Agent Interactions, 2003.
463. Vasanth Sarathy, Jason R. Wilson, Thomas Arnold, and Matthias Scheutz. Enabling Basic Normative HRI in a Cognitive
Robotic Architecture. In 2nd Workshop on Cognitive Architectures for Social Human-Robot Interac, 2016.
464. Eric L. Sauser, Brenna D. Argall, Giorgio Metta, and Aude G. Billard. Iterative learning of grasp adaptation through
human corrections. Robotics and Autonomous Systems, 60:55–71, 2012.
465. Jonathan R. Scally, Nicholas L. Cassimatis, and Hiroyuki Uchida. Worlds as a unifying element of knowledge representation.
Biologically Inspired Cognitive Architectures, 1:14–22, 2012.
466. Samer Schaat, Klaus Doblhammer, Alexander Wendt, Friedrich Gelbard, Lukas Herret, and Dietmar Bruckner. A
psychoanalytically-inspired motivational and emotional system for autonomous agents. Industrial Electronics Society,
IECON 2013-39th Annual Conference, pages 6648–6653, 2013.
467. Samer Schaat, Alexander Wendt, and Dietmar Bruckner. A multi-criteria exemplar model for holistic categorization in
autonomous agents. Industrial Electronics Society, IECON 2013-39th Annual Conference of the IEEE, pages 6642–6647,
2013.
468. Samer Schaat, Alexander Wendt, Matthias Jakubec, Friedrich Gelbard, Lukas Herret, and Dietmar Dietrich. ARS: An
AGI agent architecture. Lecture Notes in Computer Science, 8598:155–164, 2014.
469. Samer Schaat, Alexander Wendt, Stefan Kollmann, Friedrich Gelbard, and Matthias Jakubec. Interdisciplinary Devel-
opment and Evaluation of Cognitive Architectures Exemplified with the SiMA Approach. In EuroAsianPacific Joint
Conference on Cognitive Science, 2015.
470. Paul Schermerhorn, James Kramer, Timothy Brick, David Anderson, Aaron Dingler, and Matthias Scheutz. DIARC: A
Testbed for Natural Human-Robot Interactions. In Proceedings of AAAI 2006 Robot Workshop, pages 1972–1973, 2006.
471. M. Scheutz, J. McRaven, and Gy. Cserey. Fast, reliable, adaptive, bimodal people tracking for indoor environments. In
Proceedings of IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), pages 1347–1352, 2004.
A Review of 40 Years in Cognitive Architecture Research Core Cognitive Abilities and Practical Applications 69

472. Matthias Scheutz, Jack Harris, and Paul Schermerhorn. Systematic Integration of Cognitive and Robotic Architectures.
Advances in Cognitive Systems, 2:277–296, 2013.
473. Matthias Scheutz, Evan Krause, and Sepideh Sadeghi. An Embodied Real-Time Model of Language-Guided Incremental
Visual Search. In Proceedings of the 36th annual meeting of the Cognitive Science Society, pages 1365–1370, 2014.
474. Matthias Scheutz and Paul Schermerhorn. Affective Goal and Task Selection for Social Robots. Handbook of Research on
Synthetic Emotions and Sociable Robotics: New Applications in Affective Computing and Artificial Intelligence, page 74,
2009.
475. Matthias Scheutz, Paul Schermerhorn, James Kramer, and David Anderson. First steps toward natural human-like HRI.
Autonomous Robots, 22(4):411–423, 2007.
476. Marvin R. G. Schiller and Fernand R. Gobet. A comparison between cognitive and AI models of blackjack strategy
learning. Lecture Notes in Computer Science, pages 143–155, 2012.
477. Craig Schlenoff, Raj Madhavan, Jim Albus, Elena Messina, Tony Barbera, and Stephen Balakirsky. Fusing Disparate
Information Within the 4D / RCS Architecture. In Proceedings of the 7th International Conference on Information
Fusion, 2005.
478. Tobias Schroder and Paul Thagard. Priming: Constraint Satisfaction and Interactive Competition. Social Cognition,
32:152–167, 2014.
479. Rachael D. Seidler, Youngbin Kwak, Brett W. Fling, and Jessica A. Bernard. Neurocognitive Mechanisms of Error-Based
Motor Learning. Advances in Experimental Medicine and Biology, 782:39–60, 2013.
480. O.G. Selfridge. Pandemonium: a paradigm for learning in mechanisation of thought processes. In Proceedings of a
Symposium Held at the National Physical Laboratory, 1958.
481. Anil K. Seth, Jeffrey L. McKinstry, Gerald M. Edelman, and Jeffrey L. Krichmar. Visual binding through reentrant
connectivity and dynamic synchronization in a brain-based device. Cerebral Cortex, 14(11):1185–1199, 2004.
482. Daniel Shapiro, Pat Langley, and Ross Shachter. Using Background Knowledge to Speed Reinforcement Learning in
Physical Agents. In Proceedings of the 5th International Conference on Autonomous Agents, pages 254–261, 2001.
483. S. C. Shapiro, Josephine Anstey, D. E. Pape, T. D. Nayak, Michael Kandefer, and O. Telhan. MGLAIR Agents in a
Virtual Reality Drama. CSE Technical Report 2005-08, 2005.
484. Stuart C. Shapiro and Jonathan P. Bona. The GLAIR Cognitive Architecture. International Journal of Machine Con-
sciousness, 2(2):307–332, 2010.
485. Stuart C. Shapiro and Michael Kandefer. A SNePS Approach to the Wumpus World Agent or Cassie Meets the Wumpus.
In IJCAI-05 Workshop on Nonmonotonic Reasoning, Action, and Change (NRAC’05): Working Notes, 2005.
486. Lokendra Shastri. Types and Quantifiers in SHRUTI - a connectionist model of rapid reasoning and relational processing.
In International Workshop on Hybrid Neural Systems, 1998.
487. Lokendra Shastri. SHRUTI: A neurally motivated architecture for rapid, scalable inference. In Studies in Computational
Intelligence, pages 183–203. Springer Berlin Heidelberg, 2007.
488. Shinsuke Shimojo and Ladan Shams. Sensory modalities are not separate modalities: plasticity and interactions. Curr.
Opin. Neurobiol., 11(4):505–509, 2001.
489. Mohan Shiwali and John E. Laird. Learning to play Mario. Tech. Rep. CCA-TR-2009-03, 2009.
490. Kristoffer Sjoo, Hendrik Zender, Patric Jensfelt, Geert-Jan M. Kruijff, Andrzej Pronobis, Nick Hawes, and Michael Brenner.
The Explorer System. In Cognitive Systems. 2010.
491. Nady Slam, Wenjun Wang, Guixiang Xue, and Pei Wang. A framework with reasoning capabilities for crisis response
decision-support systems. Engineering Applications of Artificial Intelligence, 46:346–353, 2015.
492. Aaron Sloman. The Cognition and Affect project: Architectures, architecture-schemas, and the new science of mind.
Technical Report, 2003.
493. Ryan Small and Clare Bates Congdon. Agent Smith: Towards an evolutionary rule-based agent for interactive dynamic
games. In 2009 IEEE Congress on Evolutionary Computation, CEC 2009, pages 660–666, 2009.
494. Richard Ll. Smith, Fernand Gobet, and Peter C. R. Lane. An Investigation into the Effect of Ageing on Expert Memory
with CHREST. In Proceedings of the United Kingdom Workshop On Computational Intelligence, 2007.
495. Scott D. G. Smith, Richard Escobedo, Michael Anderson, and Thomas P. Caudell. A deployed engineering design retrieval
system using neural networks. IEEE Transactions on Neural Networks, 8(4):847–51, 1997.
496. Larry R Squire. Declarative and Nondeclarative Memory: Multiple Brain Systems Supporting Learning. Journal of
Cognitive Neurosciencecience, 4(3):232–243, 1992.
497. Janusz A. Starzyk and James T. Graham. MLECOG: Motivated Learning Embodied Cognitive Architecture. IEEE
Systems Journal. Special Issue on Human-Like Intelligence and Robotics, PP(99), 2015.
498. Terrence C. Stewart, Peter Blouw, and Chris Eliasmith. Explorations in Distributed Recurrent Biological Parsing. In
International Conference on Cognitive Modelling, 2015.
499. Terrence C. Stewart and Chris Eliasmith. Parsing Sequentially Presented Commands in a Large-Scale Biologically Realistic
Brain Model. In Proceedings of the 35th Annual Conference of the Cognitive Science Society, pages 3460–3467, 2013.
500. Terrence C. Stewart and Chris Eliasmith. Large-Scale Synthesis of Functional Spiking Neural Circuits. Proceedings of the
IEEE, 102(5):881–898, 2014.
70 Iuliia Kotseruba, John K. Tsotsos

501. Arthur Still and Mark D’Inverno. A History of Creativity for Future AI Research. In Proc. 7th Comput. Creat. Conf.,
pages 152–159, 2016.
502. Dustin Stokes and Stephen Biggs. The dominance of the visual. In D. Stokes, M. Matthen, and S. Biggs, editors, Perception
and its Modalities, pages 1–35. Oxford University Press, 2014.
503. Svorad Stolc and Ivan Bajla. Application of the computational intelligence network based on hierarchical temporal
memory to face recognition. In Proceedings of the 10th IASTED International Conference on Artificial Intelligence and
Applications (AIA), pages 185–192, 2010.
504. Ron Sun. Hybrid Connectionist-Symbolic Modules. AI Magazine, 17(2):99–103, 1996.
505. Ron Sun. Desiderata for cognitive architectures. Philosophical Psychology, 17(3):341–373, 2004.
506. Ron Sun. The importance of cognitive architectures: An analysis based on CLARION. Journal of Experimental and
Theoretical Artificial Intelligence, 19:159–193, 2007.
507. Ron Sun. Memory systems within a cognitive architecture. New Ideas in Psychology, 30(2):227–240, 2012.
508. Ron Sun. Anatomy of the Mind: Exploring Psychological Mechanisms and Processes with the Clarion Cognitive Archi-
tecture. Oxford University Press, 2016.
509. Ron Sun and Lawrence A. Bookman, editors. Computational architectures integrating neural and symbolic processes: A
perspective on the state of the art. Springer Science & Business Media, 1994.
510. Ron Sun and Pierson Fleischer. A Cognitive Social Simulation of Tribal Survival Strategies: The Importance of Cognitive
and Motivational Factors, volume 12. 2012.
511. Ron Sun and Sebastien Helie. Accounting for creativity within a psychologically realistic cognitive architecture. Compu-
tational Creativity Research: Towards Creative Machines, 7:3–36, 2015.
512. Ron Sun, Edward Merrill, and Todd Peterson. A bottom-up model of skill learning. Proceedings of 20th Cognitive Science
Society Conference, pages 1037–1042, 1998.
513. Ron Sun, Todd Peterson, and Edward Merrill. A Hybrid Architecture for Situated Learning of Reactive Sequential Decision
Making. Applied Intelligence, 11:109–127, 1999.
514. Ron Sun and Nick Wilson. Motivational Processes Within the Perception-Action Cycle. Springer New York, New York,
NY, 2011.
515. Ron Sun, Nick Wilson, and Michael Lynch. Emotion: A Unified Mechanistic Interpretation from a Cognitive Architecture.
Cognitive Computation, 8(1):1–14, 2016.
516. Ron Sun, Nick Wilson, and Robert Mathews. Accounting for Certain Mental Disorders Within a Comprehensive Cognitive
Architecture. In Proceedings of International Joint Conference on Neural Networks, 2011.
517. Ron Sun and Xi Zhang. Top-Down versus Bottom-Up Learning in Skill Acquisition. In Proceedings of the 24th Annual
Conference of the Cognitive Science Society, 2002.
518. Ron Sun and Xi Zhang. Accessibility versus Action-Centeredness in the Representation of Cognitive Skills. In Proceedings
of the Fifth International Conference on Cognitive Modeling, 2003.
519. Ron Sun, Xi Zhang, and Robert Mathews. Modeling meta-cognition in a cognitive architecture. Cognitive Systems
Research, 7(4):327–338, 2006.
520. Botond Szatmary, Jason Fleischer, Donald Hutson, Douglas Moore, James Snook, Gerald M. Edelman, and Jeffrey Krich-
mar. A Segway-based human-robot soccer team. IEEE International Conference on Robotics and Automation, 2006.
521. Niels A. Taatgen. A model of individual differences in skill acquisition in the Kanfer-Ackerman air traffic control task.
Cognitive Systems Research, 3(1):103–112, 2002.
522. Yaniv Taigman, Marc Aurelio Ranzato, Tel Aviv, and Menlo Park. DeepFace: Closing the Gap to Human-Level Perfor-
mance in Face Verification. In CVPR, 2014.
523. Guy Taylor and Lin Padgham. An Intelligent Believable Agent Environment. AAAI Technical Report WS-96-03, 1996.
524. Gheorghe Tecuci, Mihai Boicu, Michael Bowman, Dorin Marcu, Ping Shyr, and Cristina Cascaval. An Experiment in
Agent Teaching by Subject Matter Experts. International Journal of Human-Computer Studies, 53(4):583–610, 2000.
525. Gheorghe Tecuci, Mihai Boicu, Thomas Hajduk, Dorin Marcu, Marcel Barbulescu, Cristina Boicu, and Vu Le. A tool for
training and assistance in emergency response planning. In Proceedings of the Annual Hawaii International Conference
on System Sciences, pages 1–10, 2007.
526. Gheorghe Tecuci, Mihai Boicu, Dorin Marcu, and David Schum. How Learning Enables Intelligence Analysts to Rapidly
Develop Practical Cognitive Assistants. In Proceedings of the 12th International Conference on Machine Learning and
Applications, pages 105–110, 2013.
527. Gheorghe Tecuci and Yves Kodratoff. Apprenticeship Learning in Imperfect Domain Theories. In Machine Learning: An
Artificial Intelligence Approach, pages 514–552. 1990.
528. Gheorghe Tecuci, Dorin Marcu, Mihai Boicu, and Vu Le. Mixed-Initiative Assumption-Based Reasoning for Complex
Decision-Making. Studies in Informatics and Control, 16(4):459–468, 2007.
529. Gheorghe Tecuci, David Schum, Mihai Boicu, Dorin Marcu, and Benjamin Hamilton. Intelligence analysis as agent-assisted
discovery of evidence, hypotheses and arguments. Smart Innovation, Systems and Technologies, 4:1–10, 2010.
530. Paul Thagard. Cognitive Architectures. In W. Frankish and W. Ramsay, editors, The Cambridge handbook of cognitive
science, pages 50–70. Cambridge University Press, Cambridge, 2012.
A Review of 40 Years in Cognitive Architecture Research Core Cognitive Abilities and Practical Applications 71

531. Robert Thibadeau, Marcel Adam Just, and Patricia A. Carpenter. A Model of the Time Course and Content of Reading.
Cognitive Science, 6:157–203, 1982.
532. Robert Thomson, Stefano Bennati, and Christian Lebiere. Extending the Influence of Contextual Information in ACT-R
using Buffer Decay. In Proceedings of the Annual Meeting of the Cognitive Science Society, 2014.
533. Kristinn Thórisson and Helgi Helgasson. Cognitive Architectures and Autonomy: A Comparative Review. Journal of
Artificial General Intelligence, 3(2):1–30, 2012.
534. Kristinn R. Thorisson. Layered modular action control for communicative humanoids. In Conference Proceedings of
Computer Animation, pages 134–143, 1997.
535. Kristinn R. Thorisson. Real-time decision making in multimodal face-to-face communication. In Proceedings of the
Interantional Conference on Autonomous Agents, pages 16–23, 1998.
536. Kristinn R. Thorisson. Mind model for multimodal communicative creatures and humanoids. Applied Artificial Intelli-
gence, 13(4-5):449–486, 1999.
537. Kristinn R. Thorisson, Olafur Gislason, Gudny Ragna Jonsdottir, and Hrafn Th. Thorisson. A Multiparty Multimodal
Architecture for Realtime Turntaking. In International Conference on Intelligent Virtual Agents, 2010.
538. John Thornton, Jolon Faichney, Michael Blumenstein, and Trevor Hine. Character Recognition Using Hierarchical Vector
Quantization and Temporal Pooling. In Proceedings of the 21st Australasian Joint Conference on Artificial Intelligence:
Advances in Artificial Intelligence, volume 5360, pages 562–572, 2008.
539. Vadim Tikhanoff, Angelo Cangelosi, and Giorgio Metta. Integration of speech and action in humanoid robots: iCub
simulation experiments. IEEE Transactions on Autonomous Mental Development, 3(1):17–29, 2011.
540. J. Gregory Trafton, Nicholas L. Cassimatis, Magdalena D. Bugajska, Derek P. Brock, Farilee E. Mintz, and Alan C. Schultz.
Enabling Effective Human Robot Interaction Using Perspective-Taking in Robots. IEEE Transactions on Systems, Man,
and Cybernetics - Part A: Systems and Humans, 35(4):460–470, 2005.
541. J. Gregory Trafton and Anthony M. Harrison. Embodied Spatial Cognition. Topics in Cognitive Science, 3:686–706, 2011.
542. J. Gregory Trafton, Laura M. Hiatt, Anthony M. Harrison, P. Tamborello, Sangeet S. Khemlani, and Alan C. Schultz.
ACT-R/E: An Embodied Cognitive Architecture for Human-Robot Interaction. Journal of Human-Robot Interaction,
2(1):30–54, 2013.
543. Lara M Triona, Amy M Masnick, and Bradley J Morris. What Does it Take to Pass the False Belief Task ? An ACT-R
Model. 72:15213, 2001.
544. Bryan Tripp and Chris Eliasmith. Function approximation in inhibitory networks. Neural Networks, 77:95–106, 2016.
545. Nishant Trivedi, Pat Langley, Paul Schermerhorn, and Matthias Scheutz. Communicating, Interpreting, and Executing
High-Level Instructions for Human-Robot Interaction. In Proceedings of AAAI Fall Symposium: Advances in Cognitive
Systems, 2011.
546. John K. Tsotsos. Analyzing vision at the complexity level. Behavioral and Brain Sciences, 13:423–469, 1990.
547. John K. Tsotsos. Image Understanding. In Encyclopedia of Artificial Intelligence, pages 641–663. 1992.
548. John K. Tsotsos. A computational perspective on visual attention. MIT Press, 2011.
549. John K. Tsotsos. Attention and Cognition: Principles to Guide Modeling. In Q. Zhao, editor, Computational and Cognitive
Neuroscience of Vision. Elsevier, 2017.
550. John K. Tsotsos, Scan M. Culhane, Winky Yan Kei Wai, Yuzhong Lai, Neal Davis, and Fernando Nuflo. Modeling visual
attention via selective tuning. Artificial Intelligence, 78(1-2):507–545, 1995.
551. John K. Tsotsos and Wouter Kruijne. Cognitive programs: Software for attention’s executive. Frontiers in Psychology,
5:1–16, 2014.
552. E. Tulving. Episodic and Semantic memory. In E. Tulving and W. Donaldson, editors, Organ. Mem., pages 382–402.
Academic Press, 1972.
553. Sherman W. Tyler, Christian Neukom, Michael Logan, and Jay Shively. The MIDAS human performance model. In
Proceedings of the Human Factors and Ergonomics Society, pages 320–324, 1998.
554. Patrick Ulam, Ashok Goel, and Joshua Jones. Reflection in Action: Model-Based Self-Adaptation in Game Playing Agents.
In Challenges in Game Artificial Intelligence: Papers from the AAAI Workshop., 2004.
555. Baris Ulutas, Erdem Erdemir, and Kazuhiko Kawamura. Application of a hybrid controller with non-contact impedance
to a humanoid robot. In Proceedings of the IEEE 10th International Workshop on Variable Structure Systems, pages
378–383, 2008.
556. Volkan Ustun, Paul S. Rosenbloom, Julia Kim, and Lingshan Li. Building High Fidelity Human Behavior Models in the
Sigma Cognitive Rchitecture. In L. Yilmaz, W. K. V. Chan, I. Moon, T. M. K. Roeder, C. Macal, and M. D. Rossetti,
editors, Proceedings of the 2015 Winter Simulation Conference, 2015.
557. Niels Van Hoorn, Julian Togelius, and Jürgen Schmidhuber. Hierarchical controller learning in a first-person shooter. In
2009 IEEE Symposium on Computational Intelligence and Games, pages 294–301, 2009.
558. K. VanLehn. Discovering problem solving strategies: What humans do and machines don’t (yet). In Proceedings of the
Sixth International Workshop on Machine Learning., pages 215–217, 1989.
559. K. VanLehn, W. Ball, and B. Kowalski. Non-Lifo Execution of Cognitive Procedures. Cognitive Science, 13(3):415–465,
1989.
72 Iuliia Kotseruba, John K. Tsotsos

560. Kurt VanLehn and William Ball. Goal Reconstruction: How Teton Blends Situated Action and Planned Action. Technical
report, Department of Computer Science and Psychology, Carnegie Mellon University, 1989.
561. Kurt Vanlehn, William Ball, and Bernadette Kowalski. Explanation-based learning of correctness: Towards a model of
the self-explanation effect. In Proceedings of the 12th Annual Conference of the Cognitive Science Society, 1990.
562. Sashank Varma. A computational model of Tower of Hanoi problem solving. PhD Thesis, 2006.
563. Manuela Veloso. PRODILOGY/ANALOGY: Analogical reasoning in general problem solving. In Topics in Case-Based
Reasoning, 1993.
564. Manuela M. Veloso and Jim Blythe. Linkability: Examining Causal Link Commitments in Partial-order Planning. In
Proceedings of the Second International Conference on Artificial Intelligence Planning Systems, 1994.
565. Manuela M. Veloso, Martha E. Pollack, and Michael T. Cox. Rationale-Based Monitoring for Planning in Dynamic
Environments. In AIPS 1998 Proceedings, pages 171–180, 1998.
566. Steven Vere and Timothy Bickmore. A basic agent. Computational Intelligence, 6(1):41–60, 1990.
567. Steven A. Vere. Organization of the basic agent. ACM SIGART Bulletin, 2(4):164–168, 1991.
568. David Vernon, Giorgio Metta, and Giulio Sandini. A Survey of Artificial Cognitive Systems: Implictions for the Au-
tonomous Development of Mental Capbilities in Computational Agents. IEEE Transactions on Evolutionary Computation,
pages 1–30, 2007.
569. David Vernon, Claes von Hofsten, and Luciano Fadiga. The iCub Cognitive Architecture. In A Roadmap for Cognitive
Development in Humanoid Robots, pages 121–153. 2010.
570. Paul Verschure and Philipp Althaus. A real-world rational agent: Unifying old and new AI. Cognitive Science, 27(4):561–
590, 2003.
571. Y. Vinokurov, C. Lebiere, A. Szabados, S. Herd, and R. O’Reilly. Integrating top-down expectations with bottom-up
perceptual processing in a hybrid neural-symbolic architecture. Biologically Inspired Cognitive Architectures, 6:140–146,
2013.
572. Vasiliki Vouloutsi, Maria Blancas Munoz, Klaudia Grechuta, Stephane Lallee, Armin Duff, Jordi ysard Llobet Puigbo,
and Paul F. M. J. Verschure. A new biomimetic approach towards educational robotics: the Distributed Adaptive Control
of a Synthetic Tutor Assistant. 4th International Symposium on New Frontiers in human-Robot Interaction, 2015.
573. Dirk Walther, Laurent Itti, Maximilian Riesenhuber, Tomaso Poggio, and Christof Koch. Attentional Selection for Object
Recognition a Gentle Way. In International Workshop on Biologically Motivated Computer Vision, 2002.
574. Dirk B. Walther and Christof Koch. Attention in Hierarchical Models of Object Recognition. Progress in brain research,
165, 2007.
575. Di Wang, Budhitama Subagdja, Ah-hwee Tan, and GW Ng. Creating human-like autonomous players in real-time first
person shooter computer games. In Proceedings of the 21st Annual Conference on Innovative Applications of Artificial
Intelligence, pages 173–178, 2009.
576. Jian Wang, Golshah Naghdy, and Philip Ogunbona. Wavelet-based feature-adaptive adaptive resonance theory neural
network for texture identification. Journal of Electronic Imaging, 6(3):329–336, 1997.
577. Pei Wang. Rigid flexibility: The logic of intelligence, volume 34. Springer Netherlands, 2006.
578. Pei Wang. Three fundamental misconceptions of Artificial Intelligence. Journal of Experimental & Theoretical Artificial
Intelligence, 19(3):249–268, 2007.
579. Pei Wang. Non-Axiomatic Logic (NAL) Specification, 2010.
580. Pei Wang. Natural language processing by reasoning and learning. In Proceedings of the International Conference on
Artificial General Intelligence, pages 160–169, 2013.
581. Pei Wang and Patrick Hammer. Assumptions of decision-making models in AGI. International Conference on Artificial
General Intelligence, pages 197–207, 2015.
582. Pei Wang and Patrick Hammer. Issues in Temporal and Causal Inference. In Proceedings of the International Conference
on Artificial General Intelligence, 2015.
583. Yongjia Wang and John E. Laird. Integrating Semantic Memory into a Cognitive Architecture. Technical Report CCA-
TR-2006-02, 2006.
584. Carter Wendelken and Lokendra Shastri. Connectionist mechanisms for cognitive control. Neurocomputing, 65-66:663–672,
2005.
585. John Carter Wendelken. SHRUTI-agent: A structured connectionist architecture for reasoning and decision-making. PhD
Thesis, 2003.
586. Juyang Weng. A theory for mentally developing robots. In Proceedings of the 2nd International Conference on Develop-
ment and Learning, pages 131–140, 2002.
587. Juyang Weng and Wey Shiuan Hwang. From neural networks to the brain: Autonomous mental development. IEEE
Computational Intelligence Magazine, 1(3):15–31, 2006.
588. Juyang Weng and Wey Shiuan Hwang. Incremental hierarchical discriminant regression. IEEE Transactions on Neural
Networks, 18(2):397–415, 2007.
589. Juyang Weng, Yong-Beom Lee, and Colin H. Evans. The Developmental Approach to Multimedia Speech Learning. In
Proceedings., of the IEEE International Conference on Acoustics, Speech, and Signal Processing, 1999.
A Review of 40 Years in Cognitive Architecture Research Core Cognitive Abilities and Practical Applications 73

590. Juyang Weng and Matthew Luciw. Online learning for attention, recognition, and tracking by a single developmental
framework. In Proceedings of the Conference on Computer Vision and Pattern Recognition, 2010.
591. Juyang Weng and Yilu Zhang. Developmental Robots - A New Paradigm. In Proceedings of the Second International
Workshop on Epigenetic Robotics Modeling Cognitive Development in Robotic Systems, volume 94, pages 163–174, 2002.
592. Dirk Wentura and Klaus Rothermund. Priming is not priming is not priming. Social Cognition, 32:47–67, 2014.
593. Stefan Wermter. Hybrid Approaches to Neural Network-based Language Processing. Technical Report TR-97-030, 1997.
594. Stefan Wermter and Ron Sun, editors. Hybrid Neural Systems. Springer-Verlag Berlin Heidelberg, New York, 2000.
595. Christopher D. Wickens, Jason S. Mccarley, Amy L. Alexander, Lisa C. Thomas, Michael Ambinder, and Sam Zheng.
Attention-Situation Awareness (A-SA) Model of Pilot Error. Human performance modeling in aviation, pages 213–239,
2008.
596. Tom Williams, Gordon Briggs, Bradley Oosterveld, and Matthias Scheutz. Going Beyond Literal Command-Based In-
structions: Extending Robotic Natural Language Interaction Capabilities. AAAI, pages 1387–1393, 2015.
597. Tom Williams and Matthias Scheutz. A Framework for Resolving Open-World Referential Expressions in Distributed
Heterogeneous Knowledge Bases. In Proceedings of the Thirtieth AAAI Conference on Artificial Intelligence, 2016.
598. Jason R. Wilson, Kenneth D. Forbus, and Matthew D. McLure. Am I really scared? A multi-phase computational model
of emotions. Proceedings of the Second Annual Conference on Advances in Cognitive Systems, pages 289–304, 2013.
599. Jason R. Wilson, Evan Krause, Morgan Rivers, and Matthias Scheutz. Analogical Generalization of Actions from Single
Exemplars in a Robotic Architecture. In Proceedings of the 2016 International Conference on Autonomous Agents &
Multiagent Systems, 2016.
600. Jason R. Wilson and Matthias Scheutz. Analogical Generalization of Activities from Single Demonstration. In Proceedings
of Ibero-American Conference on Artificial Intelligence, pages 637–648, 2014.
601. Nicholas R. Wilson and Ron Sun. Coping with Bullying: A Computational Emotio -Theoretic Account. In CogSci, 2014.
602. Nicholas R. Wilson, Ron Sun, and Robert C. Mathews. A motivationally-based simulation of performance degradation
under pressure. Neural Networks, 22(5-6):502–508, 2009.
603. Nicholas R. Wilson, Ron Sun, and Robert C. Mathews. A Motivationally Based Computational Interpretation of Social
Anxiety Induced Stereotype Bias. In Proceedings of the 2010 cognitive science society conference, pages 1750–1755, 2010.
604. Samuel Wintermute. An Overview of Spatial Processing in Soar/SVS Investigator. Technical Report CCA-TR-2009-01,
2009.
605. Samuel Wintermute. Imagery in cognitive architecture: Representation and control at multiple levels of abstraction.
Cognitive Systems Research, 19-20:1–29, 2012.
606. Michael T. Wolf, Christopher Assad, Yoshiaki Kuwata, Andrew Howard, Hrand Aghazarian, David Zhu, Thomas Lu,
Ashitey Trebi-Ollennu, and Terry Huntsberger. 360-Degree visual detection and target tracking on an autonomous surface
vehicle. Journal of Field Robotics, 27(6):819–833, 2010.
607. Jeremy M Wolfe. Guided Search 2.0 A revised model of visual search. Psychonomic bulletin & review, 1(2):202–238, 1994.
608. C. Wong, D. Kortenkamp, and M. Speich. A mobile robot that recognizes people. In Proceedings of 7th IEEE International
Conference on Tools with Artificial Intelligence, 1995.
609. Dean Wyatte, Seth Herd, Brian Mingus, and Randall O’Reilly. The role of competitive inhibition and top-down feedback
in binding during object recognition. Frontiers in Psychology, 3, 2012.
610. Xuhong Xiao, Gee Wah Ng, Yuan Sin Tan, and Yeo Ye Chuan. Scene parsing and fusion-based continuous traversable
region formation. In C. Jawahar and Shan S., editors, Computer Vision - ACCV 2014 Workshops, 2015.
611. John Yen, Michael McNeese, Tracy Mullen, David Hall, Xiaocong Fan, and Peng Liu. RPD-based Hypothesis Reasoning
for Cyber Situation Awareness. In Cyber Situational Awareness, pages 39–49. 2010.
612. Chen Yu, Matthias Scheutz, and Paul Schermerhorn. Investigating multimodal real-time patterns of joint attention in an
HRI word learning task. In Human-Robot Interaction (HRI), 2010 5th ACM/IEEE International Conference on, pages
309–316, 2010.
613. Wayne Zachary, Thomas Santarelli, J. Ryder, and James Stokes. Developing a multi-tasking cognitive agent using the
COGNET/iGEN integrative architecture. Technical report, 2000.
614. Wayne W. Zachary, Jean-Christoph Le Mentec, and Joan M. Ryder. Interface Agents in Complex Systems. Human
Interaction with Complex Systems, 14(1):260–264, 2016.
615. Wayne W Zachary, Joan M Ryder, James H Hicinbothom, Janis A Cannon-Bowers, and Eduardo Salas. Cognitive task
analysis and modeling of decision making in complex environments. In J. Cannon-Bowers and E. Salas, editors, Making
decisions under stress: Implications for individual and team training., pages 315–344. Washington, DC, 1998.
616. Wayne W. Zachary, Allen L. Zaklad, James H. Hicinbothom, Joan M. Ryder, and Janine A. Purcell. COGNET Reprezen-
tation of Tactical Decision-Making in Anti-Air Warfare. In Proceedings of the Human Factors and Ergonomics Society
37th Annual Meeting, pages 1112–1116, 1993.
617. N. Zhang, J. Weng, and Z. Zhang. A developing sensory mapping for robots. In Proceedings 2nd International Conference
on Development and Learning. ICDL 2002, pages 13–20, 2002.
618. Yilu Zhang and Juyang Weng. Task transfer by a developmental robot. IEEE Transactions on Evolutionary Computation,
11(2):226–248, 2007.
74 Iuliia Kotseruba, John K. Tsotsos

619. Wen Zhuo, Zhiguo Cao, Yueming Qin, Zhenghong Yu, and Yang Xiao. Image classification using HTM cortical learning
algorithms. In Proceedings of the 21st International Conference on Pattern Recognition (ICPR), pages 2452–2455, 2012.
620. Sharon Zmigrod and Bernhard Hommel. Feature Integration across Multimodal Perception and Action: A Review. Mul-
tisensory Research, 26:143–157, 2013.