Analogy and Qualitative

Representations in the
Companion Cognitive
Architecture
Kenneth D. Forbus, Thomas Hinrichs, Northwestern University
Abstract: The Companion cognitive architecture is aimed at reaching human-level AI by creating software social
organisms, systems that interact with people using natural modalities, working and learning over extended periods of
time as collaborators rather than tools. Our two central hypotheses about how to achieve this are (1) analogical
reasoning and learning are central to cognition, and (2) qualitative representations provide a level of description that
facilitates reasoning, learning, and communication. This paper discusses the evidence we have gathered supporting
these hypotheses from our experiments with the Companion architecture. Although we are far from our ultimate
goals, these experiments provide strong evidence for the utility of analogy and QR across a range of tasks. We also
discuss three lessons learned and highlight three important open problems for cognitive systems research more
broadly.

Introduction
Every cognitive architecture starts with a set of theoretical commitments. We have argued (Forbus, 2016) that human-
level artificial intelligences will be built by creating sufficiently smart software social organisms. By that we mean systems
capable of interacting with people using natural modalities, operating and learning over extended periods of time, as
apprentices and collaborators, instead of as tools. Just as we cannot directly access the internal representations of the
people and animals we work with, cognitive systems should be able to work with us on our terms. But how does one
create such systems? We have two core hypotheses, inspired by research in cognitive science:
1. Analogical reasoning and learning are central to human cognition. There is evidence that processes described by
Gentner’s (1983) structure-mapping theory of analogy and similarity operate throughout human cognition,
including visual perception (Sagi et al. 2012), reasoning and decision-making (Markman & Medin, 2002), and
conceptual change (Gentner et al. 1997).
2. Qualitative representations (QR) are a key building block of human conceptual structure. Continuous phenomena
and systems permeate our environment and our ways of thinking about it. This includes the physical world,
where qualitative representations have a long track record of providing human-level reasoning and
performance (Forbus, 2014), but also in social reasoning (e.g. degrees of blame (Tomai & Forbus, 2007)).
Qualitative representations carve up continuous phenomena into symbolic descriptions that serve as a bridge
between perception and cognition, facilitate everyday reasoning and communication, and help ground expert
reasoning.
The focus of the Companion cognitive architecture is on higher-order cognition: Conceptual reasoning and learning,
and learning via interactions with others, by contrast with architectures that have focused on skill learning (e.g. ACT-
R (Anderson, 2009) and SOAR (Laird, 2012)). In Newell’s (1990) time-scale decomposition of cognitive phenomena,
conceptual reasoning and learning occur in what are called the Rational and Social bands1, unlike many architectures
which start with Newell’s Cognitive band (Laird et al., this issue). Thus we approximate subsystems whose operations

1 In models of some processes, e.g. analogical matching, we capture phenomena at the Cognitive band as well.

1
occur at faster time-scales, using methods whose outputs have reasonable cognitive fidelity, although we are not
making theoretical claims about them. For example, in Companions constraint checking and simple inferences are
carried out via a logic-based truth-maintenance system (Forbus & de Kleer, 1994). Similarly, the Companions
architecture is implemented as a multi-agent system, capable of running on a single laptop or distributed across many
cluster nodes, depending on the task. We suspect that many data-parallel operations and a number of coarse-grained
parallel operations are important for creating robust software organisms, but this current organization is motivated
more by engineering concerns. By contrast, we spend considerable effort on sketch understanding and natural
language understanding, since interaction via natural modalities is a core concern.
Cognitive systems are knowledge-rich (e.g. McShane, this issue). The Companion architecture uses the Cyc2 ontology
and knowledge base (KB) contents, plus our own extensions including additional linguistic resources, representations
for visual/spatial reasoning, and the Hierarchical Task Network (HTN) plans that drive Companion operations. This
is in sharp distinction with current practice in machine learning, where the goal is that systems for every task must be
learned from scratch. The ML approach can be fine for specific applications, but it means that many independently-
learnable factors in a complex task must be learned at the same time. This intermingling leads to requiring far more
training data than people need, and reduces transfer of learned knowledge from one task to another. The cognitive
systems approach strives for cumulative learning, where an architecture that is general-purpose can learn to do one
task and successfully applies that knowledge to learn other tasks.
The rest of this paper summarizes the evidence provided by Companion experiments for our core hypotheses. We
start by summarizing the ideas underlying our hypotheses, briefly introducing structure-mapping and our models of
its processes for analogy followed by the key aspects of qualitative reasoning. Then we describe the evidence that
analogy and qualitative representations play a variety of roles in cognition. We close with some lessons learned and
open problems.

Analogical Reasoning and Learning

Gentner’s (1983) structure-mapping theory proposed that analogy involves the construction of mappings between two
structured, relational representations. These mappings contain correspondences (i.e. what goes with what), candidate
inferences, (i.e. what can be projected from one description to the other, based on the correspondences), and a score
indicating the overall quality of the match. Typically, the base is the description about which more is known, and the
target is the description about which one is trying to reason about, and hence inferences are made from base to target
by default. Inferences in the other direction, reverse candidate inferences, can also be drawn, which is important for
difference detection, discussed more below.
There is ample evidence for the psychological plausibility of structure-mapping, out of Gentner’s lab and others
(Forbus, 2001). This evidence includes some implications that may surprise AI researchers. Here are three such
findings:
1. The same computations that people use for analogy are also used for everyday similarity (Markman &
Gentner, 1993). Most AI researchers model similarity as the dot product of feature vectors or other distributed
representations. But those approaches are not compatible with psychological evidence that indicate, even for
visual stimuli, relational representations are important (Goldstone et al. 1991).
2. Differences are closely tied to similarity (Markman & Gentner, 1996). There is an interesting dissociation
between difference detection and naming a difference: It is faster to detect that things are different when they
are very different, but faster to name a difference when they are very similar (Sagi et al. 2012). This falls
directly out of our computational model of analogical matching below.
3. Analogical comparison, especially within-domain comparison, can happen unconsciously. For example,
people reading a story can be pushed between two different interpretations of it, based on a different story
that they read previously (Day & Gentner, 2007). Moreover, they report being sure that the second story was
complete in itself, and nothing from the previous story influenced them. This implies, for example, that in
any of the System 1/System 2 models (e.g. Kahneman 2011), analogy and similarity are actually used in both
systems.
Our computational models of analogical processes have been designed both as cognitive models and as performance
systems. That is, each of them has been used to explain (and sometimes predict) psychological findings, while also
being used in AI systems whose sole goal was achieving some new functionality. These models are
 SME: The Structure-Mapping Engine (Forbus et al. 2016) models analogical mapping. It computes up to three
mappings, using a greedy algorithm to operate in polynomial time. It can also operate incrementally,
extending mappings as new facts are added to the base or target, which can be useful in problem-solving and
language understanding.

2 www.cyc.com

2
  MAC/FAC (Many Are Called/Few Are Chosen): MAC/FAC (Forbus et al. 1995) models analogical retrieval,
using two map/reduce stages. The MAC stage uses a non-structured vector representation, automatically
derived from structured representations, such that their dot products provide an estimate of how SME will
score the original structural representations. The best, and up to two more if sufficiently close, are passed to
the FAC stage, which uses SME in parallel to match the original structured representations. The MAC stage
provides scalability, while the FAC stage provides the precision and inferences needed to support reasoning.
 SAGE (Sequential Analogical Generalization Engine): SAGE (McLure et al. 2015) models analogical
generalization. For each concept, it maintains a generalization pool that is incrementally updated as new
examples arrive. The pool contains generalizations, which are structured descriptions with a probability
attached to each statement, and unassimilated examples. For each new example, the most similar item from
the pool is retrieved via MAC/FAC. If sufficiently similar, they are assimilated. The assimilation process
involves updating the probabilities for each statement, based on the contents of the match and the differences,
and replacing any non-identical entities with skolems. Thus commonalities become highlighted and
accidental information becomes deprecated. Note that logical variables are not introduced, since
generalizations can still be applied to new situations via analogy.
In computational terms, we think of these as an analogy stack: An organization of analogical processing for cognitive
systems that we believe has wide applicability. In Companions, analogical operations are more primitive than
backchaining. Rule-based reasoning and logical deduction are used for encoding, case construction, constraint
checking, and simple inferences. The architecture also includes an AND/OR graph problem solver and an HTN
planner, but all of these freely use the analogical processes above in their operations.

Qualitative Representations
People have robust mental models of the world that help them operate in the world, e.g. we cook, navigate, and bond
with others. A common concern in these models is how to deal with continuous properties and phenomena. In
reasoning about the physical world, we must deal with quantities like heat, temperature, and mass, and arrange for
processes like flows and phase changes to cook. Spatial factors are key in many physical situations, such as finding
safe ground during a flood. In social reasoning, we think about quantities like degree of blame, how strong a friendship
is, and the relative morality of choices we must make. Even in metacognition, we estimate how difficult problems are,
and decide whether to carry on or give up on a tough problem. Qualitative representations provide more abstract
descriptions of continuous parameters and phenomena, which are more in line with data actually available to
organisms. We may not be able to estimate accurately how fast something is moving, but we can easily distinguish
between rising and falling, for example. People routinely come to reasonable conclusions with very little information,
and often do so rapidly. Combining qualitative representations with analogy provides an explanation for this (Forbus
& Gentner, 1997).
Broadly, qualitative representations can be decomposed into those concerned with quantities and those concerned with
space. Quantities are an important special case because many continuous parameters are one dimensional, and this
imposes additional constraints on reasoning. We use qualitative process theory (QP theory, Forbus, 1984) as our
account of quantities and continuous processes. Spatial qualitative representations are typically grounded in metric
representations (i.e. the Metric Diagram/Place Vocabulary model (Forbus et al. 1991)), which provide the functional
equivalent of visual processing to support spatial reasoning. Once places are constructed, often purely qualitative
reasoning can be done on them, via spatial calculi (Cohn & Renz 2008). Our current model of visual and spatial
reasoning, CogSketch (Forbus et al. 2011), automatically produces cognitively plausible visual and spatial
representations from vector inputs (e.g. digital ink, copy/paste from PowerPoint), and is integrated into the Companion
architecture.

Some Roles of Analogy and Qualitative Representations in Cognition

We provide evidence for our hypotheses by examining a range of tasks, showing how analogy and qualitative
representations have been used to enable Companions to perform them3.

Reasoning
Textbook problem solving involves solving the kinds of problems that are posed to students learning science and
engineering. In experiments conducted with the aid of the Educational Testing Service, we showed that a Companion,

3Here we only stick to models that use the entire Companion cognitive architecture. Other models indicate that these
same ideas are relevant in modeling visual analogies, including geometric analogies (Lovett et al. 2009b), oddity tasks
(Lovett & Forbus, 2011), and Ravens’ Progressive Matrices (Lovett & Forbus, 2017).

3
using representations generated by ETS and Cycorp, could learn to transfer knowledge via analogy to new problems
across six different within-domain conditions (Klenk & Forbus, 2009). It operated by using MAC/FAC to retrieve
worked solutions, which were at the level of detail found in textbooks, rather than the internals of the reasoning system,
which neither ETS nor Cycorp had access to. Equations to solve new problems were projected as candidate inferences
onto the new problem, with their relevance ascertained by overlapping event structure between the new and old
problem. This included solving larger, more complex problems by combining analogies involving two solutions from
simpler problems. Qualitative representations played three roles in this reasoning: First, some problems themselves
are qualitative in their nature, such as asking whether something will move at all in a given situation, and if so, will its
velocity and acceleration be constant or increasing. The second role was expressing the conditions under which an
equation is applicable. The third role was translating real-world conditions into parameter values (e.g. at the top of its
trajectory, the velocity of a rising object is zero).
The role of qualitative representations for reasoning found in earlier stages of science education is even larger. Before
high school, children are not asked to formulate equations, but to reason qualitatively, draw conclusions from graphs
and tables, and tie knowledge learned in class to their everyday experience. An analysis of fourth-grade science tests
suggests that qualitative representations and reasoning are sufficient for at least 29% of their questions (Crouse &
Forbus, 2016), and even more when qualitative spatial and visual representations are taken into account.

Decision-Making
An example of decision-making is MoralDM (Deghani et al. 2008), which models the influence of utilitarian concerns
and sacred values on human-decision-making. An example of a sacred value is not taking an action that will directly
result in someone’s death, even if it would save more lives (see Scheutz, this issue, for examples). MoralDM modeled
the impact of sacred values via a qualitative order of magnitude representation, so that when a sacred value was
involved, the utilitarian differences seemed negligible compared to violating that value. While it had some rules for
detecting that a situation should invoke a sacred value, it also used analogy to apply such judgements from similar
situations. Experiments indicate that increasing the size of case library improves its decision, as measured against
human judgments in independent experiments (Deghani et al. 2008), and using analogical generalization over cases
improved performance further (Blass & Forbus, 2015).

Commonsense Reasoning
Analogical reasoning may be at the heart of commonsense reasoning: Similar prior experiences can be used to make
predictions but also provide explanations for observed behavior (Forbus, 2015). Moreover, even partial explanations
relating aspects of a situation can be projected via candidate inferences, without a complete and correct logically
quantified domain theory. For example, questions from a test of plausible reasoning can be answered by chaining
multiple small analogies, similar to how rules are chained in deductive reasoning (Blass & Forbus, 2016).

Understanding Language
Our research indicates that analogy can play at least four roles in natural language understanding. First, analogies are
often explicitly stated in language, particularly in instructional materials. These need to be detected so that subsequent
processing will gather up the base and target and the appropriate conclusions can be drawn. Analogical dialogue acts
(Barbella & Forbus, 2011) extend communication act theories with additional operations inspired by structure-
mapping, enabling a Companion to learn from such analogies expressed in natural language, as measured by improved
query-answering performance. Second, for integrating knowledge learned by reading, it can be useful to carve up
material in a complex text into cases. These cases can be used for answering comparison questions directly (Barbella
& Forbus, 2015), and for subsequent processing (e.g. rumination (Forbus et al. 2007)). Third, analogy can also be used
to learn evidence for semantic disambiguation, by accumulating cases about how particular word/word sense pairings
were used in the linguistic context (including lexical, syntactic, and semantic information), and honed by SAGE to
suggest how to disambiguate future texts (Barbella & Forbus, 2013). Finally, analogy can be used to learn new linguistic
constructions (McFate & Forbus, 2016a), e.g. understanding sentences like “Sue crutched Joe the apple so he wouldn’t
starve.”
Our evidence to date suggests that QP theory provides a formalism for natural language semantics concerning
continuous phenomena. Qualitative representations capture the level of detail found in linguistic descriptions, and
mappings can be established from FrameNet representations (McFate & Forbus, 2016b) to qualitative process theory.
Moreover, qualitative representations can be learned by reading, with the knowledge produced improving strategy
game performance (McFate et al. 2014).

Domain Learning
Learning from instruction often involves multiple modalities, e.g. text and diagrams. Analogy can be used to fuse
cross-modal information, by using SME to align the representations constructed for each modality and projecting

4
information across them. For example, Lockwood (Lockwood & Forbus, 2009) showed that semi-automatic language
processing of a simplified English chapter of a book on basic machines, combined with sketches for the diagrams,
enabled a system to learn enough to correctly answer 12 out of 15 questions from the back of the chapter. Chang (2016)
has shown that these ideas can work when the language processing is fully automatic, and that by using a corpus of
instructional analogies aimed at middle-school instruction, expressed in simplified English and sketches, a Companion
learned enough to correctly answer 11 out of 14 questions from the New York Regent’s examination on the topics
covered.
Companions have also been used to do cross-domain analogical learning, both in physical domains (Klenk & Forbus,
2009) and in games (Hinrichs & Forbus, 2011). For cross-domain learning, building up persistent mappings across
multiple analogies, essentially making a translation table between domains, was important. In cross-domain game
learning, SME was used to build up a meta-mapping between predicates in the two domains, which then enabled the
rest of the domain knowledge to be imported.
Qualitative representations provide a useful level of causal knowledge for learning the dynamics of complex domains
(Hinrichs & Forbus, 2012a), and can be used as a high-level representation for goals and strategies (Hinrichs & Forbus,
2014). Qualitative reasoning supports reflective reasoning by treating goal activation levels as continuous quantities in
order to represent mental states (Hinrichs & Forbus, 2016). For example, in reasoning about a strategy game such as
Freeciv4, the learned qualitative model of domain dynamics enables identifying tradeoffs, and high-level strategies
such as first expanding one’s empire, then building up its economy, can be expressed by processes which are carried
forward by a player’s actions.

Conceptual Change
One of the well-documented phenomena in cognitive science is conceptual change, i.e. how people’s mental models
develop over long spans of time (e.g. diSessa et al. 2004; Ioannides et al. 2002). Catalogs of misconceptions have been
created for a variety of phenomena, and in some cases, trajectories that learners take in mastering these concepts have
been analyzed and documented. Friedman’s assembled coherence theory of conceptual change takes as its starting point
qualitative, compositional model fragments, but initially tied to specific experiences. Analogical retrieval is used to
determine which model fragments are retrieved and combined to create a model for a new situation. This theory
explains both why mental models often seem very specific to
particular classes of situations and yet, within those classes, are Figure 1: An example of a forced-choice
relentlessly consistent. Model fragments are constructed via SAGE task. Which goes with the one on top?
and a set of transformation heuristics, and applied via MAC/FAC.
Friedman’s Companion-based TIMBER system has successfully
modeled the trajectories of intuitive force models (Friedman &
Forbus, 2010) using sketched behaviors, transitions between
misconceptions when learning about the human circulatory system
(Friedman & Forbus, 2011), misconceptions about why there are
seasons (Friedman et al. 2011), and debugging misconceptions about
the day/night cycle by analogy (Friedman et al. 2012).

Cognitive State Sensing

Our mental life is governed in part by the ability to sense our internal
state: In doing a forced-choice task, like the one in Figure 1, often one choice “looks right” and we take it, even if we
cannot always articulate why this is so. If it isn’t obvious, we have to think about it more. We believe that the structural
evaluation score computed by SME is a signal used in these computations. In simulating this task (Kandaswamy et al.
2014), if there is a strong difference between SME’s scores for the two comparison, the highest score is chosen. But if
not, re-representation techniques are used to modify the encodings of the stimuli (which are automatically generated
via CogSketch) to attempt to discriminate between them. This model has captures the phenomena in several
psychological experiments.
Another kind of cognitive state sensing, in our view, are emotions. One property of emotions is that they are often
rapid, but not always specific. We (Wilson et al. 2013) suspect this is because appraisal information used in generating
emotions is stored with memories, and retrieved via analogy to produce a rapid response. This response is
subsequently modified by slower cognitive processes, and eventually when new memories are consolidated, the
appraisals are adjusted based on the outcome of the situation. Hence a problem that might have initially seemed scary,
if mastered, gets stored as something doable. With this multi-phase model of emotions, a Companion learning to solve
problems was able to perform more effectively, and with emotional dynamics consistent with the literature.

4 http://www.freeciv.org/

5
Classification
The analogy stack described above can be used for classification, by doing analogical retrieval over the union of
generalization pools representing the possible concepts of interest. The label associated with the pool where the best
reminding came from serves as the identification of the concept, with the score providing an estimate of the quality of
the classification and the correspondences of the mapping providing an explanation as to why that classification is
appropriate. The candidate inferences provide further surmises about the new example. This method has been used
to learn geospatial concepts in a strategy game (e.g. isthmus, bay, etc.), achieving 62% accuracy. By extending SAGE
with automatic near-miss detection (summarized below), accuracy rose to 77% (McLure et al. 2015).

A meta-analogy
Analogy is more conservative than explanation-based learning, which generalizes after only one example, but requires
many fewer examples than feature-based statistical learning systems – sometimes orders of magnitude fewer (e.g.
Kuehne et al. 2000). Can we do even better by somehow combining ideas of structure-mapping with ideas from
machine learning? Here is a meta-analogy that suggests a family of approaches: structure-mapping is to relational
representations what dot product is to feature vectors. This suggests that any machine learning technique defined on
feature vectors originally might be adapted to a hybrid method by replacing dot product with SME. For example, by
combining a structured form of logistic regression with SAGE on the link plausibility/knowledge base completion task,
Liang & Forbus (2015) showed that this approach can yield state of the art results, with several orders of magnitude
fewer examples, while also providing material for explanations. Another hybrid has been shown to produce state of
the art performance on the Microsoft Paraphrase task (Liang et al. 2016). The advantage of using task-specific training
is that similarity weights can be tuned to specific tasks, and learned negative weights (which SME does not produce)
can improve discrimination. The drawback is that the representations learned, unless the weights are stored separately,
become task-specific instead of general-purpose. We believe this is a very useful approach that can be applied to many
techniques, e.g. using SME as a kernel for SVMs. But we know of no evidence that this is psychologically plausible.

Some Lessons Learned

We have learned a lot through these and other experiments. We summarize the three most important lessons here.
First, the models of structure-mapping processes turn out to be surprisingly robust. We have only made two extensions
in tackling the span of tasks above. The first are interim generalizations, i.e. a small set of generalization pools that exist
in working memory, in contrast with normal generalization pools, which are held in long-term memory. Interim
generalizations were motivated by both the rapid learning of children in experiments and by the unconscious story
intrusion effects in adults. In the experiments above, they are used in learning forced-choice tasks as a simple form of
re-representation, i.e. when an interim generalization is retrieved for an item, non-overlapping aspects of the item are
filtered out. The second are near-misses. Winston’s (1970) concept of near-miss required a teacher to know the
machine’s single concept, and formulate an example that differed in only one way. McLure’s observation (McLure et
al. 2015) was that retrieving an item from a different generalization pool over items in the intended label is, when the
two labels are mutually exclusive, a near-miss. He extended SAGE to construct positive and negative probabilistic
hypotheses based on near-misses, which are added to the retrieval criteria for the generalization pool to improve
discriminability.
Second, type-level qualitative representations (Hinrichs & Forbus, 2012b) are important for scalability and for flexible
reasoning. In traditional qualitative reasoning, a system might start with a schematic or scenario description, and then
formulate and use models to reason about it. By contrast, consider playing a strategy game like Freeciv, where
constructing a civilization is the goal. A successful civilization involves a dozen or more cities, each with many
parameters, and even more units. Formulating and maintaining explicit propositional qualitative models for each city,
much less building a complete snapshot of the game state, can be prohibitively expensive. Moreover, it is often
necessary to reason about planned units and cities that don’t exist yet. Type-level representations make these problems
simpler, although still not easy. They are amenable to analogy (because they are quantifier-free) and also useful in
natural language semantics: The QP frame representations we generate are neutral with respect to whether they are
specific or generic, leaving this decision to later semantic processing, which has the contextual information needed to
determine this.
Finally, we have found the Cyc ontology and KB contents to be extremely valuable. We developed our own reasoning
engine because we needed source code access and, for us, analogy operations are central. The breadth and depth of
the Cyc ontology has saved us countless hours, by letting us use existing representations instead of building our own.
Even when our requirements diverge from Cyc’s ontology, small additions usually suffice to bridge the gap. For

6
example, we treat quantities as fluents, as opposed to specific values, so a single logical function suffices to translate5.
Moreover, we have found microtheories to be an excellent way to impose modularity on KB organization and support
more flexible reasoning. Every operation is done with respect to a logical environment, e.g. a microtheory and those it
inherits from, thereby controlling what knowledge is used in a computation. Alternatives and qualitative states are
represented as microtheories, as are subsketches within a sketch. In our reasoning engine, microtheories are also cases,
which makes analogical reasoning even more tightly integrated.

Discussion and Open Problems

Our experience with Companions to date has provided strong evidence for our core hypotheses. Analogical processing
and qualitative representations are useful for textbook problem solving, moral decision-making, and commonsense
reasoning. Analogy can play multiple roles in natural language understanding and visual problem solving, and
qualitative representations provide a natural level of expressiveness for semantics and reasoning. Together they can
successfully model a variety of phenomena in conceptual change, and provide a means of fusing information across
modalities. They can be used inside cognitive systems for cognitive state sensing, to improve problem solving. They
can be used for classification tasks. And there is now evidence that, by combining analogical processing with ideas
from standard machine learning, state of the art performance on tasks of interest to the machine learning community
can be attained, sometimes with orders of magnitude fewer examples and increased explainability.
To be sure, many open problems remain. We outline three here. The first is scaling. How much knowledge might be
needed to achieve human-level AI? Aside from the initial endowment of knowledge, how much experience might be
needed? A back of the envelope calculation (Forbus et al. in press) suggests that a lower bound of at least 2 million
generalization pools, built up by at least 45 million examples, would be needed to approximate some reasonable
fraction of what a college graduate knows. For Companions, we will need to exploit the data-parallel capabilities built
into our analogical models via hardware, to reach this scale and maintain real-time performance. The second is that
being able to use natural modalities to interact with people, in a fluent and adaptive manner, is a key bottleneck for
progress. While for Companions simplified English syntax has been very productive, we would prefer them to be able
to puzzle through more complex syntax. Our current hope is that learning linguistic constructions via analogy will
enable this, but that is far from clear. The final issue is longevity, i.e. creating cognitive systems that successfully
perform while learning over weeks, months, and years at a time. Most cognitive architectures, including ours, are fired
up for experiments, the experiments are run, and then they are shut down. Achieving robust learning while
maintaining accurate performance, even with people inspecting the system’s internals, is quite challenging (Mitchell
et al. 2015). Enabling Companions to build up solid internal models of their own reasoning and learning processes
seems to be the most promising approach to us. We think that the notions of attention and control that Bello and
Bridewell (this issue) explore will be an important part of the solution to longevity.

Acknowledgements
We thank the Office of Naval Research, the Air Force Office of Scientific Research, the National Science Foundation,
and the Defense Advanced Research Projects Agency for their support via multiple programs. Statements and opinions
expressed may not reflect the position or policy of the United States Government, and no official endorsement should
be inferred. David Barbella, Maria Chang, Morteza Dehghani, Scott Friedman, Matthew Klenk, Kate Lockwood,
Andrew Lovett, Joe Blass, Max Crouse, Subu Kandaswamy, Chen Liang, Clifton McFate, Matthew McLure, Irina
Rabkina, and Madeline Usher contributed heavily to this research.

References
Anderson, J. R. (2009) How can the Human Mind Occur in the Physical Universe? Oxford University Press.
Barbella, D. and Forbus, K. (2011). Analogical Dialogue Acts: Supporting Learning by Reading Analogies in
Instructional Texts. Proceedings of the Twenty-Fifth AAAI Conference on Artificial Intelligence (AAAI 2011), San Francisco,
CA.
Barbella, D. and Forbus, K. (2013). Analogical Word Sense Disambiguation. Advances in Cognitive Systems, 2:297–315.
Barbella, D. & Forbus, K. (2016) Exploiting Connectivity for Case Construction in Learning by Reading. Advances in
Cognitive Systems 4:169–186.
Bello, P. & Bridewell, W. (2017) There is no Agency Without Attention. AI Magazine, this issue.
Blass, J. and Forbus, K.D. (2015). Moral Decision-Making by Analogy: Generalizations vs. Exemplars. Proceedings of the
Twenty-Ninth AAAI Conference on Artificial Intelligence, Austin, Texas.

5The logical function QPQuantityFn we introduced has as its domain Cyc continuous quantities and has as its range
QP theory fluents.

7
Blass, J. A. & Forbus, K. D. (2016) Modeling Commonsense Reasoning via Analogical Chaining: A Preliminary Report.
Proceedings of the 38th Annual Meeting of the Cognitive Science Society, Philadelphia, PA, August.
Cohn, A., and Renz, J., 2008. Qualitative Spatial Representation and Reasoning, in: F. van Hermelen, V. Lifschitz, B.
Porter, eds., Handbook of Knowledge Representation, Elsevier, 551-596.
Crouse, M., & Forbus, K. (2016) Elementary school science as a Cognitive System Domain: How Much Qualitative
Reasoning is Required? Proceedings of the 4th Annual Conference on Advances in Cognitive Systems.
Day, S. & Gentner, D. (2007). Nonintentional analogical inference in text comprehension. Memory and Cognition, 35, 39–
49.
Dehghani, M., Tomai, E., Forbus, K., Klenk, M. (2008). An Integrated Reasoning Approach to Moral Decision-Making.
Proceedings of the Twenty-Third AAAI Conference on Artificial Intelligence (AAAI). Chicago, IL.
diSessa, A., Gillespie, N., Esterly, J. (2004) Coherence versus fragmentation in the development of the concept of force.
Cognitive Science, 28,843-900.
Forbus, K. (1984). Qualitative process theory. Artificial Intelligence, 24, 85–168.
Forbus, K. (2001). Exploring analogy in the large. In Gentner, D., Holyoak, K., and Kokinov, B. (Eds.) Analogy:
Perspectives from Cognitive Science. MIT Press.
Forbus, K., (2014) Qualitative Reasoning. In Teofilo Gonzalez, Jorge Diaz-Herrera, Allen Tucker (Eds.) Computing
Handbook, Third Edition: Computer Science and Software Engineering. CRC Press.
Forbus, K. (2015). Analogical Abduction and Prediction: Their Impact on Deception. AAAI Fall Symposium on Deceptive
and Counter-Deceptive Machines.
Forbus, K. (2016). Software Social Organisms: Implications for Measuring AI Progress. AI Magazine, 37(1):85–90.
Forbus, K. & de Kleer, J. (1994) Building Problem Solvers. MIT Press.
Forbus, K., Nielsen, P. and Faltings, B. (1991) Qualitative Spatial Reasoning: The CLOCK Project, Artificial Intelligence,
51 (1-3), October, 1991.
Forbus, K., Riesbeck, C., Birnbaum, L., Livingston, K., Sharma, A., and Ureel, L. (2007). Integrating Natural Language,
Knowledge Representation and Reasoning, and Analogical Processing to Learn by Reading. Proceedings of AAAI-07:
Twenty-Second Conference on Artificial Intelligence, Vancouver, BC.
Forbus, K., Ferguson, R., Lovett, A., & Gentner, D. (2016). Extending SME to Handle Large-Scale Cognitive Modeling.
Cognitive Science, DOI:10.1111/cogs.12377.
Forbus, K., Gentner, D. and Law, K. 1995. MAC/FAC: A model of Similarity-based Retrieval. Cognitive Science, 19(2),
April-June, pp 141–205.
Forbus, K., Klenk, M., and Hinrichs, T. 2009. Companion Cognitive Systems: Design Goals and Lessons Learned So
Far. IEEE Intelligent Systems, 24(4), 36–46.
Forbus, K., Liang, C., & Rabkina, I. (in press) Representation and Computation in Cognitive Models. Topics in Cognitive
Science.
Forbus, K., Usher, J., Lovett, A., Lockwood, K., and Wetzel, J. (2011). CogSketch: Sketch understanding for Cognitive
Science Research and for Education. Topics in Cognitive Science, 3(4), pp 648–666.
Friedman, S., & Forbus, K. (2008). Learning Causal Models via Progressive Alignment and Qualitative Modeling: A
Simulation. Proceedings of CogSci 2008.
Friedman, S. E. & Forbus, K. (2010). An integrated systems approach to explanation-based conceptual change. In
Proceedings of the Twenty-Fourth AAAI Conference on Artificial Intelligence. Atlanta, GA.
Friedman, S., & Forbus, K. (2011). Repairing incorrect knowledge with model formulation and metareasoning.
Proceedings of the 22nd International Joint Conference on Artificial Intelligence. Barcelona, Spain.
Friedman, S. E., Forbus, K. D. and Sherin, B. (2011). How do the seasons change? Creating & revising explanations via
model formulation & metareasoning. Proceedings of the 25th International Workshop on Qualitative Reasoning. Barcelona,
Spain.
Friedman, S., Barbella, D., and Forbus, K. (2012). Revising Domain Knowledge with Cross-Domain Analogy. Advances
in Cognitive Systems, 2, 13–24.
Gentner, D. (1983). Structure-mapping: A theoretical framework for analogy. Cognitive Science, 7, 155–170.
Gentner, D., Brem, S., Ferguson, R. W., Markman, A. B., Levidow, B. B., Wolff, P., & Forbus, K. D. (1997). Analogical
reasoning and conceptual change: A case study of Johannes Kepler. The Journal of the Learning Sciences, 6(1), 3–40.
Goldstone, R. L., Medin, D. L., & Gentner, D. (1991). Relational similarity and the nonindependence of features in
similarity judgments. Cognitive Psychology, 23, 222–264.
Hinrichs, T. and Forbus, K. (2011). Transfer Learning Through Analogy in Games. AI Magazine, 32(1), 72–83.
Hinrichs, T. and Forbus, K. (2012a). Learning Qualitative Models by Demonstration. Proceedings of the 26th AAAI
Conference on Artificial Intelligence, 207–213, July.
Hinrichs, T. and Forbus, K. (2012b). Toward Higher-Order Qualitative Representations. Proceedings of the Twenty-
sixth International Workshop on Qualitative Reasoning. Los Angeles CA, July.

8
Hinrichs, T., & Forbus, K. (2014) X goes First: Teaching Simple Games through Multimodal Interaction. Advances in
Cognitive Systems, 3:31–46.
Hinrichs, T. and Forbus, K. (2016). Qualitative Models for Strategic Planning. Advances in Cognitive Systems, 4:75–92.
Ioannides, C., & Vosniadou, S. (2002). The changing meanings of force. Cognitive Science Quarterly, 2:5-61.
Kahneman, D. (2011) Thinking, Fast and Slow. Allen Lane.
Kandaswamy, S., Forbus, K., and Gentner, D. (2014). Modeling Learning via Progressive Alignment using Interim
Generalizations. Proceedings of the Cognitive Science Society.
Klenk, M. and Forbus, K. (2009). Analogical Model Formulation for AP Physics Problems. Artificial Intelligence, 173(18),
1615–1638. doi:10.1016/j.artint.2009.09.003
Klenk, M. & Forbus, K. (2013). Exploiting persistent mappings in cross-domain analogical learning of physical domains
Artificial Intelligence, 195:398–417.
Kuehne, S., Gentner, D. and Forbus, K. (2000). Modeling infant learning via symbolic structural alignment. Proceedings
of CogSci 2000, August.
Laird, J. (2012). The SOAR cognitive architecture. Cambridge: MIT Press
Laird, J., Lebiere, C., & Rosenbloom, P. (2017) A Standard Model of the Mind: Toward a Common Computational
Framework Across Artificial Intelligence, Cognitive Science, Neuroscience, and Robotics. AI Magazine, this issue.
Liang, C. and Forbus, K. (2015). Learning Plausible Inferences from Semantic Web Knowledge by Combining
Analogical Generalization with Structured Logistic Regression. Proceedings of the Twenty-Ninth AAAI Conference on
Artificial Intelligence, Austin, Texas.
Liang, C., Paritosh, P., Rajendran, V., and Forbus, K. D. (2016). Learning Paraphrase Identification with Structural
Alignment. Proceedings of IJCAI-2016
Lockwood, K. and Forbus, K. (2009). Multimodal knowledge capture from text and diagrams. Proceedings of KCAP-
2009.
Lovett, A., Gentner, D., Forbus, K., & Sagi, E. (2009a). Using analogical mapping to simulate time-course phenomena
in perceptual similarity. Cognitive Systems Research, 10. 216–228.
Lovett, A., Tomai, E., Forbus, K., & Usher, J. (2009b). Solving geometric analogy problems through two-stage analogical
mapping. Cognitive Science, 33(7), 1192–1231.
Lovett, A., & Forbus, K. (2017) Modeling visual problem solving as analogical reasoning. Psychological Review, 124(1),
pp. 60-90.
Lovett, A., & Forbus, K. (2011). Cultural commonalities and differences in spatial problem-solving: A computational
analysis. Cognition, 121(2), 281–287.
Markman, A.B., & Gentner, D. (1993) Structural Alignment during Similarity Comparisons. Cognitive Psychology,
25,431–467.
Markman, A. B., & Gentner, D. (1996). Commonalities and differences in similarity comparisons. Memory & Cognition,
24(2), 235–249.
Markman, A. B., & Medin, D. L. (2002). Decision making. Stevens' Handbook of Experimental Psychology
McFate, C.J., Forbus, K. and Hinrichs, T. (2014). Using Narrative Function to Extract Qualitative Information from
Natural Language Texts. Proceedings of the Twenty-Eighth AAAI Conference on Artificial Intelligence, Québec City,
Québec, Canada.
McFate, C., and Forbus, K. (2016a). Analogical Generalization and Retrieval for Denominal Verb Interpretation.
Proceedings of the 38th Annual Meeting of the Cognitive Science Society, Philadelphia, PA, August.
McFate, C., and Forbus, K. (2016b). An Analysis of Frame Semantics of Continuous Processes. Proceedings of the 38th
Annual Meeting of the Cognitive Science Society, Philadelphia, PA, August.
McLure, M.D., Friedman S.E. and Forbus, K.D. (2015). Extending Analogical Generalization with Near-Misses.
Proceedings of the Twenty-Ninth AAAI Conference on Artificial Intelligence, Austin, Texas.
McShane, M. (2017) Natural Language Understanding (NLU, not NLP) in Cognitive Systems. AI Magazine, this issue.
Mitchell, T., Cohen, W., Hruschka, E., Talukdar, P., Betteridge, J., Carlson, A., Dalvi, B., Gardner, M., Kisiel, B.,
Krishnamurthy, J., Lao, N., Mazaitis, K., Mohamed, T., Nakashole, N., Platanios, E., Ritter, A., Samadi, M., Settles, B.,
Wang, R., Wijaya, D., Gupta, A., Chen, X., Saparov, A., Greaves, M., Welling, J. (2015). Never-Ending Learning.
Proceedings of AAAI 2015.
Newell, A. (1990) Unified Theories of Cognition. Harvard University Press.
Sagi, E., Gentner, D., & Lovett, A. (2012). What difference reveals about similarity. Cognitive Science, 36(6), 1019–1050.
Scheutz, M. (2017) The Case for Explicit Ethical Agents. AI Magazine, this issue.
Tomai, E. and Forbus, K. (2007). Plenty of Blame to Go Around: A Qualitative Approach to Attribution of Moral
Responsibility. In Proceedings of Qualitative Reasoning Workshop 2007, Aberystwyth, U.K.
Wilson, J., Forbus, K., McLure, M. (2013) Am I Really Scared? A Multiphase Computational Model of Emotions.
Proceedings of Advances in Cognitive Systems.
Winston, P. H., 1970. Learning structural descriptions from examples. Ph.D. diss., MIT, Cambridge, MA

9
Kenneth D. Forbus is the Walter P. Murphy Professor of Computer Science and Professor of Education at Northwestern University.
His research interests include qualitative reasoning, analogical reasoning and learning, natural langauge understanding, sketch
understanding, and cognitive architecture.

Thomas Hinrichs is a Research Associate Professor of Computer Science at Northwestern University. His research interests include
commonsense reasoning, cognitive architectures and machine learning.

10