Threaded Cognition Published

Psychological Review Copyright 2008 by the American Psychological Association
2008, Vol. 115, No. 1, 101–130 0033-295X/08/$12.00 DOI: 10.1037/0033-295X.115.1.101
Threaded Cognition: An Integrated Theory of Concurrent Multitasking
Dario D. Salvucci Niels A. Taatgen

Drexel University Carnegie Mellon University and University of Groningen
The authors propose the idea of threaded cognition, an integrated theory of concurrent multitasking—that
is, performing 2 or more tasks at once. Threaded cognition posits that streams of thought can be
represented as threads of processing coordinated by a serial procedural resource and executed across
other available resources (e.g., perceptual and motor resources). The theory specifies a parsimonious
mechanism that allows for concurrent execution, resource acquisition, and resolution of resource
conflicts, without the need for specialized executive processes. By instantiating this mechanism as a
computational model, threaded cognition provides explicit predictions of how multitasking behavior can
result in interference, or lack thereof, for a given set of tasks. The authors illustrate the theory in model
simulations of several representative domains ranging from simple laboratory tasks such as dual-choice
tasks to complex real-world domains such as driving and driver distraction.
Keywords: executive control, dual-task performance, perceptual refractory period, driving, cognitive
architectures
One of the most impressive aspects of the human cognitive cessing bottlenecks in various stages of perceptual, cognition, and
system is the ability to manage and execute multiple concurrent motor processing (e.g., Broadbent, 1958; Keele, 1973; Pashler,
tasks. Concurrent multitasking (which we often refer to as simply 1994). Such studies and theories emphasized the role of various
multitasking) is a ubiquitous phenomenon in our daily lives, from processes as representative of distinct resources in which bottle-
work to leisure to everyday routine. In some situations, multitask- necks may or may not appear (depending on particulars of the task
ing can seem nearly effortless (e.g., walking and talking); for other domain). Further work codified these efforts into related but dis-
situations, it can seem extremely difficult if not impossible (e.g., tinct theories of processing on multiple resources. Navon and
reading and listening to two distinct sentences); for still others, Gopher’s (1979) multiple-resource theory specified how tasks
multitasking performance may depend heavily on the individual using separate resources may proceed simultaneously without in-
and/or the environment (e.g., singing while playing an instrument terference, but in the presence of resource conflicts, the required
or dialing a phone while driving). This space of possibilities raises
resource can allocate part of its processing time to each task.
an enormous challenge in understanding, on the one hand, the
Wickens (2002; see also Wickens, 1984) posited an alternate
human system’s astonishing capacity for multitasking and, on the
theory of multiple resources that accounts for performance using
other hand, the sometimes severe limitations on multitasking per-
formance. In this article, we propose a new theory called threaded the four resource dimensions of processing stages, perceptual
cognition that provides a theoretical and computational framework modalities, visual channels, and processing codes. Both efforts
for understanding, modeling, and predicting performance during attempted to unify the many empirical findings related to perfor-
the concurrent execution of arbitrary tasks. mance bottlenecks into more comprehensive frameworks with
Meyer and Kieras (1997a) provided an in-depth historical back- which to reason about and quantify resource needs and potential
ground of the study of multitasking performance, dating back task interference.
several decades (e.g., Craik, 1948; Telford, 1931; Welford, 1952) In relating resources and bottlenecks to the larger issue of
and covering the vast literature concerning the potential for pro- cognitive control, several theorists have posited the need for ex-
ecutive processes that manage and schedule individual tasks. Nor-
man and Shallice (1986) described a two-component system that
performs contention scheduling for basic routine action but re-
Dario D. Salvucci, Department of Computer Science, Drexel University;
quires a supervisory attentional system for higher level control.
Niels A. Taatgen, Department of Psychology, Carnegie Mellon University
and Department of Artificial Intelligence, University of Groningen, Gro- Baddeley (1986) defined a central executive responsible for man-
ningen, the Netherlands. agement and selection of processing routines in the context of
This work was supported by Office of Naval Research (ONR) Grant working memory. Such theories have been influential in broaden-
N00014-03-1-0036 to Dario D. Salvucci and ONR Grant N00014-06-1- ing our understanding of multitasking by considering the processes
0055 to Niels A. Taatgen. We thank Erik Altmann, Erik Reichle, Stefani underlying multitasking behavior and how these processes contend
Nellen, Yelena Kushleyeva, and John Anderson for their many comments
for and utilize available resources. At the same time, there has long
and suggestions; Ron Chong for sharing his implementation of the tracking
task; and Susan Chipman for her generous support.
been a recognized desire for further specification of the mecha-
Correspondence concerning this article should be addressed to Dario D. nisms that embody these executive processes, in order to, as has
Salvucci, Department of Computer Science, Drexel University, 3141 been said, banish the “homunculus” of cognitive control (see, e.g.,
Chestnut Street, Philadelphia, PA 19104. E-mail: salvucci@cs.drexel.edu Altmann, 2003; Logan, 2003; Monsell & Driver, 2000).
101
102 SALVUCCI AND TAATGEN
With this desire in mind, many theorists have recently turned to does not require supervisory or executive processes. Instead, this
computational modeling as a methodology for providing rigorous ability is best represented by a general, domain-independent, par-
specifications of both executive and task processes. One body of simonious mechanism that allows for concurrent processing and
work has examined multitasking performance in smaller scale provides basic resource conflict resolution (see Liu, Feyen, &
laboratory tasks, modeling and accounting for various effects of, Tsimhoni, 2005). Although higher order planning may still arise in
for example, stimulus or response modality and task difficulty in deliberate reasoning about one’s own multitasking behavior, it is
dual-choice tasks (e.g., Byrne & Anderson, 2001; Meyer & Kieras, this basic ability that allows us to perform multiple concurrent
1997b; Taatgen, 2005). Another body of work has modeled mul- tasks as often arises in everyday behavior.
titasking behavior as it occurs in particular complex task do- In this article, we propose a theory of threaded cognition that
mains—for example, models of fighter piloting (Jones et al., provides both a conceptual theory and an associated computational
1999), driving (Salvucci, 2005, 2006), air-traffic control (Taatgen framework for general domain-independent multitasking.
& Lee, 2003), and similar domains (see Gluck & Pew, 2005, and Threaded cognition posits that cognition maintains an active set of
Gray, 2007, for further examples). Still other work has focused on task goals that result in threads of processing across a set of
computational models of executive control in routine activities available resources. Specifically, it posits that threads are coordi-
(Cooper & Shallice, 2000), which could in principle serve as a nated by a serial procedural resource that combines inputs from
general multitasking mechanism; however, these models currently other resources (e.g., perceptual and motor resources) and initiates
provide no theory to guide the specification or learning of higher new processing on these resources. The theory specifies how
level schemas required to resolve resource conflicts and thus have threads acquire and release resources and how conflicts with
not been applied to modeling concurrent multitasking. Overall, this respect to acquisition of the central procedural resource are re-
range of work has demonstrated the many benefits of a computa- solved. As such, the theory claims that concurrent multitasking
tional modeling approach, including a finer grained analysis of does not require supervisory executive processes; instead, concur-
behavior over time and the mapping of behavior to quantitative rent multitasking emerges from the interaction of autonomous
measures that can be compared directly with human performance process threads in conjunction with a straightforward mechanism
data. However, these efforts have almost exclusively used domain- for resource acquisition and conflict resolution. The computational
specific mechanisms of multitasking fine-tuned for a single task instantiation of this mechanism, based in the ACT-R (Adaptive
domain, making it difficult to generalize the models and mecha- Control of Thought-Rational) cognitive architecture (Anderson et
nisms to other (even closely related) task domains. al., 2004), allows multiple models of arbitrary tasks to be executed
In summary, more general theoretical treatments such as those together, generating immediate predictions of the resulting multi-
of Norman and Shallice (1986) and Baddeley (1986) do not tasking behavior. As such, threaded cognition provides a domain-
provide sufficient detail for instantiation as computational mech- independent theory and framework for understanding, represent-
anisms, whereas current computational models do not provide ing, and predicting multitasking performance.
sufficient breadth to generalize across task domains. We wish to
achieve the best of both approaches—that is, to develop a general, Threaded Cognition
domain-independent model of multitasking as a fully specified
computational mechanism. In one such effort, Kieras, Meyer, The core idea of threaded cognition is that multitasking behavior
Ballas, and Lauber (2000) (see also Kieras, 2007) developed can be represented as the execution of multiple task threads,
models in their EPIC (Executive-Process Interactive Control) cog- coordinated by a serial cognitive processor and distributed across
nitive architecture (Meyer & Kieras, 1997a) to explore the chal- multiple processing resources. In this section, we formalize our
lenges of unifying domain-specific customized executives into a notion of threaded cognition and threads in two sets of core
single domain-independent general executive. They developed assumptions. The first set of assumptions lays the theoretical
customized executives for several tasks and, for one task in par- groundwork by presenting an underlying framework for single-
ticular (tracking and choice, detailed later), also developed two task performance. The second set of assumptions extends the
versions of a general executive and compared their predictions single-task assumptions into our broader theory of threaded cog-
with those of the customized executive. However, they found that nition and multitask performance. In all, threaded cognition rep-
neither version of their general executive could adequately account resents a consolidated effort to combine our two recent individual
for human behavior in this task and resorted to a customized approaches to multitasking (Salvucci, 2005; Taatgen, 2005) into a
executive to achieve their best account of the human data. single more cohesive account.
In our view, approaches that require supervisory or executive Before delving into the details of threaded cognition, we first
processes to manage multitasking behavior have a significant wish to develop an intuitive feel for the theory by drawing an
limitation in that they do not adequately account for multitasking analogy—namely, that human cognition, particularly aspects re-
between two arbitrary tasks. For example, a person may be well lated to multitasking and executive control, behaves much like a
practiced in skills such as washing dishes or performing mental cook preparing food in a kitchen. A cook (in any variety, from a
arithmetic but may never have performed both tasks together. restaurant chef to a short-order cook to a parent cooking for a
Nevertheless, even in the first attempt, a person can perform such family) manages a number of resources in order to complete a
tasks concurrently when asked. Any approach that requires exec- stream of orders. Let us imagine a cook at a cooking station,
utive processes must account for this ability either by defining a illustrated in Figure 1a, with several available resources: an oven
general executive or by defining how specialized executives for for baking, a stove and pot for boiling, and a mixer for mixing. Let
the two tasks can be learned. We believe that humans have a basic us also imagine that orders are received on individual slips of
ability to perform multiple concurrent tasks and that this ability paper that indicate the desired dish. For a given dish, the cook must
THREADED COGNITION 103
(a)
Make Fish
Orders Make Pasta Baking Oven
Make Cake Fish
Cook:
Start boil
Mixer Batter Pasta Stove

Mixing Boil
(b)
Fish Pasta Cake
Cook: Preheat oven (waiting until cook is free)

(waiting until cook is free)
Cook: Start water to boil
Oven: Heat up to temperature Cook: Put ingredients in mixer
Stove: Heat water to boiling
Cook: Start mixer to make batter
Cook: Place fish in oven (waiting until cook is free)

Mixer: Mix cake batter
Cook: Place pasta in boiling water
Cook: Stop mixer

Stove: Boil pasta
Oven: Bake fish
Cook: Strain pasta in colander

(waiting until oven is free)
Cook: Remove fish from oven
Cook: Put batter in oven
Oven: Bake cake
Figure 1. Cooking example. a: Storyboard view with cooking resources. b: Timelines for making fish, pasta,
and cake.
execute a well-practiced sequence of steps to achieve the desired cook handles the next step for the other order. Thus, the cook can
results. For instance, to prepare fish, pasta, or a cake, a cook might often allow multiple orders to proceed in parallel, but conflicts for
execute the (grossly simplified) steps shown in Figure 1b. Note either the cook or another resource sometimes reduce parallelism
that each step generally involves the cook initiating some process and lead to delays in order processing, shown in gray in the figure.
(e.g., boiling or baking), waiting for the process to complete, and In the example, the fish order experiences no delays, the pasta
then moving on to the next step. Each step is thus associated with order experiences slight delays because of contention over the
both the desired food order and the current state of the various cook, and the cake order experiences large delays because of
resources. contention for the oven. We might even imagine that the cook
Now, imagine that the cook must work to fill multiple orders requires a cookbook for an unfamiliar order, in which case fre-
simultaneously—for example, to prepare all three dishes in the quent accesses to the cookbook would result in further delays.
figure at the same time. On the one hand, some processes can In threaded cognition, the central procedural resource can be
clearly execute in parallel—for instance, the fish can bake in the analogized to the cook, and other resources (e.g., perceptual and
oven while the pasta boils on the stove. On the other hand, several motor) can be analogized to the various cooking resources avail-
constraints limit the amount of parallelism in filling orders. First, able to the cook. The procedural resource, like the cook, collects
the cook may experience a resource conflict when two or more processing results from resources and redirects this information to
orders require the same resource, such as the fish and cake orders make new processing requests to other resources; goals, rather than
that require the oven at the same time (let us assume with different orders, guide the processing to directed action in service of a
temperatures). Second, the cook himself or herself is sometimes desired task. Conflicts can arise either when multiple tasks require
the bottleneck, specifically when processes for two or more orders the same peripheral resource (e.g., when two tasks both require
require attention from the cook at the same time—for instance, vision) or when multiple tasks require attention from the central
when the oven-preheating and the water-boiling processes end procedural resource (e.g., when auditory and visual tasks finish
simultaneously. In this case, one order must be delayed while the encoding simultaneously and require the procedural resource to
proceed). And just as a cook may require a cookbook in the initial be placed in the visual buffer and thus be available for subsequent
stages of learning, threaded cognition initially relies on memorized processing. We further define the individual resources in the
instructions but transforms a skill to a more highly proceduralized assumptions that follow.
process through learning. In essence, the interplay of resource
parallelism with potential for resource conflicts gives rise to a rich Cognitive Resources Assumption: Cognitive resources in-
array of possible multitasking scenarios that demonstrate both the clude separate procedural and declarative resources, each of
power and limitations of human multitasking. We now define the which can independently become a source of processing
theory of threaded cognition more formally in the two sets of core interference.
assumptions.
The majority of theories of dual-task performance treat cen-
tral cognition as a single resource, focusing on the ability to
Core Assumptions for Single-Task Performance translate a perceptual stimulus into a motor response (e.g.,
Pashler, 1994). In contrast, we adopt the view that cognition can
In developing our theory of multitask performance, we wish to better be characterized as two distinct resources (e.g., Anderson
start with a firm theoretical grounding for single-task performance et al., 2004; Meyer & Kieras, 1997a; cf. Laird, Newell, &
that we can then extend to multitask performance. The theoretical Rosenbloom, 1987): a declarative resource that serves as a
arguments in the following assumptions draw heavily from theo- storage memory for factual knowledge and a procedural re-
ries of resource allocation and utilization (e.g., Navon & Gopher, source that integrates information and effects new behavior.
1979), attention and control (e.g., Norman & Shallice, 1986), This distinction facilitates generalization beyond simple
production systems (e.g., Newell & Simon, 1972), and computa- stimulus–response selection to a wide range of complex tasks.
tional cognitive architectures (e.g., Meyer & Kieras, 1997a). In In addition, the distinction helps to account for a separation
addition, to make the theory concrete as a computational formal- between memory-related and procedural interference that some-
ism, we adopt the representations and mechanisms of the ACT-R times occurs in multitasking, including how practice and learn-
cognitive architecture (Anderson et al., 2004) along with various ing can lead to proceduralization and changing usage of cog-
computational treatments that have extended the architecture in nitive resources as described shortly.
ways relevant to multitasking (e.g., Byrne & Anderson, 2001;
Salvucci, 2001a; Taatgen & Lee, 2003; Taatgen, van Rijn, & Declarative Resource Assumption: Cognition’s declarative
Anderson, 2007). resource represents static knowledge as information chunks
that can be recalled (or forgotten).
Processing Resources Assumption: Human processing re-
sources include cognitive, perceptual, and motor resources. Declarative memory can be characterized as long-term storage
of information, and the declarative resource allows for access to
Our first core assumption sets the groundwork for the various this information while accounting for the effects of learning,
processing facilities, or resources (see, e.g., Navon & Gopher, decay, and retrieval failures. For our computational framework, we
1979; Norman & Bobrow, 1975), available to the human sys- adopt the ACT-R (Anderson et al., 2004) representation of this
tem. These resources can be broadly characterized as related to information in terms of chunks of knowledge, each of which
the cognitive, perceptual, and motor systems. The perceptual includes a set of attribute-value pairs such that the chunk’s type
resources acquire information from the external world and defines its attributes. For example, the chunk
include a variety of systems that facilitate visual, auditory,
tactile, olfactory, and other types of perception. The motor Three-Plus-Four
resources enable the body to perform actions in the external isa addition-fact
world, as a response to stimuli or sometimes to facilitate per- addend1 Three
ception (e.g., eye movements for visual perception or hand addend2 Four
movements for tactile perception). The cognitive resources sum Seven
process the information entering the perceptual resources and
guide further perception and motor action based on both exter- represents the knowledge of the addition fact 3 ⫹ 4 ⫽ 7. The type
nal situations and internal state. This broad assumption is, in addition-fact defines the three attributes corresponding to the
essence, nothing new but follows a long line of research on addends and sum, and the attribute values Three, Four, and Seven
processing bottlenecks (e.g., Broadbent, 1958; Pashler, 1994), are themselves chunks; thus, a set of chunks essentially forms a
theories of resource integration and interference (e.g., Wickens, network of declarative knowledge. The declarative module can
1984, 2002), and cognitive architectures (Anderson et al., 2004; receive requests to retrieve particular chunks of knowledge on the
Meyer & Kieras, 1997a). basis of a partial pattern (e.g., 3 ⫹ 4 ⫽ ?). If it succeeds in
In our computational modeling of these resources, we make the retrieving a chunk, it will place it in the retrieval buffer, the buffer
assumption that a resource has two components: a module that associated with the declarative resource. The declarative module
performs the processing related to that resource and one or more can handle only one retrieval request at a time. The chunk repre-
buffers that serve as a communication channel between the module sentation serves as the base representation for all information
and (as we describe below) the procedural resource. For example, passed between production rules and the various resource modules
a request to the visual resource to encode a visual stimulus would through the buffers—that is, resource processing requests as well
evoke processing in the visual module, and when completed, the as processing results are both represented by chunks in the asso-
result—that is, a representation of the encoded stimulus—would ciated module’s buffers.
In addition to the representation of chunks, ACT-R defines by a wide variety of theoretical and computational constructs.
several equations that govern the probability of chunk retrieval (vs. For instance, one could conceptualize the mapping in terms of
forgetting) and the latency of retrieval, both of which are depen- schemata (Norman & Shallice, 1986; Rumelhart, 1980; Rumel-
dent on chunk activation that decreases over time but increases hart, Smolensky, McClelland, & Hinton, 1986) or dynamic
with practice and use. The specifics of the memory theory are gating mechanisms (O’Reilly & Frank, 2006).
relevant to threaded cognition insofar as they predict the latency of We adopt the representation of such mappings as condition–
retrievals of a given memory chunk; we provide details as neces- action production rules, currently in use as the core component of
sary in the context of the model simulations described below. For several cognitive architectures (e.g., EPIC: Meyer & Kieras,
now, it suffices to state that the declarative resource can be 1997a; Soar: Laird, Newell, & Rosenbloom, 1987; 4CAPS: Just,
considered a source of cognitive interference independent of the Carpenter, & Varma, 1999). A production rule (which we often
procedural resource, particularly in situations in which processes refer to as simply a rule) defines a set of conditions and actions,
require frequent and potentially competing retrievals from declar- such that the conditions must be met for the rule to execute (or fire)
ative memory. the given actions. In the ACT-R formulation of a production rule,
both the conditions and actions utilize buffers for information
Perceptual and Motor Resources Assumption: The perceptual transfer: The conditions collate and test information placed in the
and motor resources allow for information acquisition from buffers by their respective modules, and if the rule fires, the
the environment and action in the environment. actions place new requests for resource processing in the buffers.
In addition to the buffers provided by the various other resources,
As mentioned earlier, the human system incorporates a variety
the system has a goal buffer. It can be considered as the procedural
of perceptual and motor resources to acquire information and act in
resource’s own buffer, which stores information about the current
the external world (respectively). Of the many possible resources,
goal of the system. Typically, production rules include a condition
our treatment focuses on the perceptual and motor resources most
for the goal buffer that matches only for goals of a particular
central to the types of tasks we wish to model: the visual and
type—for instance, a rule that concerns a choice task matches and
auditory perceptual resources and the manual motor resource.
fires only when the current goal indicates that the system is
Threaded cognition relies on the computational formulation of
attempting to perform a choice task. The goal buffer thus directs
perceptual and motor resources in ACT-R, parts of which (partic-
the production system only to rules currently relevant to the task at
ularly the motor) are derived from the EPIC cognitive architecture
hand. (This goal-directed aspect of a production system relates to
(Meyer & Kieras, 1997a). The visual resource has both a where
similar paradigms in alternative frameworks, such as the activation
system and a what system: The where system produces preattentive
of relevant action schemas and inhibition of irrelevant ones in
visual processing that finds objects in the environment on the basis
contention scheduling: Norman & Shallice, 1986; Cooper & Shal-
of spatial location and visual properties, and the what system
lice, 2000.)
identifies the object and places a declarative representation of the
In addition to matching and adding contents to the buffers, we
object in the visual buffer. An extension of this visual resource
assume that a production rule can test the status of the modules and
(Salvucci, 2001a), derived from a computational model of eye
buffers in two ways. First, a rule can test whether a module is busy
movements in reading (Reichle, Pollatsek, Fisher, & Rayner,
or free—that is, whether the module is currently performing a
1998), predicts the observable movements of the eye that corre-
resource-processing task, such as visually encoding a stimulus or
spond to the unobservable shifts of visual attention. The auditory
executing a motor movement. Second, a rule can test whether a
resource has analogous where and what systems for audible stimuli
module’s buffer is currently empty or filled: A module fills a
and places its results in the aural buffer. The motor resource
buffer with the results of information processing when completed,
assumes a two-hand manual module centered on standard interac-
and the buffer empties when a rule matches and uses this infor-
tion movements on a desktop computer (e.g., typing and mouse
mation—for example, when visual encoding completes, the visual
movement). The timing and other parameters of vision, audition,
module places the encoded information in the visual buffer, and
and motor movement are detailed in Anderson et al. (2004), along
then the buffer empties when a subsequent rule uses this informa-
with their original treatment in Meyer and Kieras (1997a).
tion. These two aspects of the modules and buffers are critical to
Procedural Resource Assumption: Cognition’s procedural re- our treatment of threaded cognition in that production rules can
source represents procedural skill as goal-directed production discriminate when a particular resource is in use: If the module is
rules. busy (processing) or the buffer is full (storing results of process-
ing), the resource can be considered in use; otherwise, it can be
The procedural resource, the central resource in our view of considered free, and thus a rule can initiate processing on the
threaded cognition, integrates and maps currently available resource if desired.
results of resource processing into new requests for further As an example, consider a scenario in which a person performs
resource processing—just as the cook in our earlier analogy a simple visual–manual choice task that involves visually encoding
transfers the end results of cooking processes (e.g., mixed a stimulus and pressing an associated response key. Let us also
batter) to initiate new processes (e.g., baking the batter into a assume that when the visual stimulus is presented on the left—for
cake). For example, in the context of a simple choice task, the instance, an O on the left side to produce the stimulus “O – –”
procedural resource may map an encoded visual stimulus and a (Schumacher et al., 2001)—the person should respond by pressing
recalled associated response into the request to perform a motor the right index finger. Such a task could be implemented with two
command. Mappings may be characterized as and implemented production rules:
Attend-stimulus central procedural resource is needed at each step to initiate new

IF the goal buffer contains a choice task resource processing, the procedural resource is not continually
and the visual resource is free and the visual buffer is empty busy but rather has periods of waiting for resource processes to
THEN issue a request to the visual resource to encode the stimulus complete (just as the cook waits for food to finish baking, frying,
Respond-to-left-stimulus etc., before proceeding); this aspect of the system is critical in our
IF the goal buffer contains a choice task upcoming specification of how threaded cognition interleaves the
and the visual buffer contains the stimulus “O – –” procedural processing for multiple tasks.
and the manual resource is free
Figure 2b shows a storyboard view illustrating how information
THEN issue a request to the manual resource to press the index finger.
passes between resources. The figure shows an overview of the
Both rules are directed by the same goal, namely the goal to resources, including processing modules (rounded rectangles) and
perform the choice task, and this goal is matched against the buffers (squared rectangles), with the procedural resource at the
current goal in the goal buffer. The first rule tests the status of the center. (Buffers for the manual and vocal resources are not shown
visual resource and issues a request to encode the stimulus (as- for simplicity, given that these resources do not produce results to
suming for simplicity that the location of the visual stimulus is be used by the procedural resource.) Each individual storyboard
known). When the visual resource processing finishes, the second diagram tracks the firing of a production rule, where arrows from
rule uses the information in the visual buffer (which also clears the the buffers to the procedural module indicate rule conditions and
buffer) and generates a motor request. arrows pointing back to the modules indicate rule actions. The
We can visualize the execution of this model in two ways. detection of the stimulus by the visual resource, combined with the
Figure 2a shows a process timeline illustrating the steps of the goal in the goal buffer, trigger the first rule that initiates the
choice task and the resources involved in each step, where time encoding process in the visual module. When encoding is com-
flows from top to bottom and the height of the box represents the pleted, the resulting representation (i.e., the encoded stimulus) is
time needed to execute that activity. As is apparent in this view, placed in the visual buffer. The representation in the visual buffer,
the process alternates between rule firings in the procedural re- again combined with the current goal, triggers the second rule that
source and processing in the peripheral (visual and manual) re- initiates the appropriate motor response. In both cases, the goal
sources. Each rule firing requires 50 ms on the procedural re- buffer guides the model by matching along with the other condi-
source, an estimate that has been derived over several decades of tions, providing top-down control of model processing along with
experimentation and is shared by the major architectures (includ- the more bottom-up triggering that arises from stimulus detection
ing EPIC and Soar). It is important to note that although this and encoding.
(a)
Visual-Manual Choice Task
Procedural: Attend stimulus
Visual: Encode stimulus
Procedural: Respond to left stimulus
Manual: Press index finger
(b)
Declarative Declarative
Goal Vis-Man Goal Vis-Man
Vocal Manual Vocal Manual

Attend-
Attend Attend- to
Respond Press
stimulus
stimulus leftstimulus
stimulus index finger
(detected) "O--"
Aural Encode Visual Aural Visual
stimulus
t = 0-50 ms t = 135-185 ms
The presence of a stimulus detected by the visual resource When visual encoding is complete, the encoded stimulus
combined with the current goal activate the first rule, which and the goal activate the second rule, which responds to
issues a request to the visual resource to encode the the leftward position of the stimulus by pressing the
full visual stimulus. index finger.
Figure 2. Single-choice task. a: Model timeline. b: Storyboard view.

Procedural Learning Assumption: When learning new tasks, When doing a single task, system resources—and the procedural
declarative task instructions are gradually transformed into resource in particular— execute in the service of a single goal. The
procedural rules that perform the task. goal, as mentioned, directs procedural processing by focusing
attention on task-relevant rules and away from task-irrelevant
A critical aspect of the dual cognitive resources lies in the rules. To generalize this view, this first assumption of threaded
gradual transformation of declarative task representations into cognition simply posits that cognition maintains a set of active
procedural representations. In particular, we adopt the theory of goals—that is, a set of goals that the system is currently trying to
production compilation (Taatgen & Lee, 2003) that provides a satisfy at the same time through concurrent multitasking. Of
computational mechanism by which task instructions encoded as course, the primary challenge that arises is how the system then
chunks change into ACT-R production rules with practice over allocates processing and balances execution among resources in
time. First, task instructions are encoded into declarative memory, order to run all goals to completion, and we address this challenge
typically by reading (visual encoding) of phrases and/or sentences in subsequent assumptions.
that specify these instructions. (In the absence of a full-fledged In terms of our computational formulation, the current ACT-R
syntactic parser, the instructions are often presented in simple architecture (Anderson et al., 2004) allows for a single goal,
albeit still realistic language.) represented as an ACT-R declarative chunk, to reside in the goal
Once instructions are stored as declarative knowledge, a general buffer and direct rule firings. We extend this formulation to allow
set of interpreter production rules retrieves each instruction and multiple goal chunks to coexist in the goal buffer, and in addition,
executes its desired actions. As these interpreter rules execute the we generalize the rule specification mechanism to allow for the
declarative instructions, the production compilation mechanism creation of multiple goals. In the current ACT-R theory, when a
begins to combine the general interpreter rules and the instructions rule specifies a new goal, this new goal replaces the current goal
into task-specific production rules. Production compilation is a in the goal buffer. For example, the rule
slow process in the sense that a new rule must typically be
relearned multiple times before it is able to compete successfully IF the goal buffer contains a choice task
and the motor response has been issued
with existing rules.
THEN set a new goal to wait for a new stimulus
Clearly this learning process has significant implications for
resource usage and thus for multitasking in general. Initially, while represents the termination of the current choice-task goal and,
task knowledge primarily lies in declarative memory, both the simultaneously, the initiation of a new waiting goal, as this latter
declarative and procedural resources experience heavy use—the goal replaces the former in the goal buffer. Our new framework
declarative for repeated retrieval of instruction chunks and the allows for multiple goals to be created in the action side of a rule.
procedural for repeated interpretation of these chunks. Later, after For instance, to direct the model to perform both the choice task
compilation of task knowledge into production rules, declarative and a secondary task, a rule could be specified as follows:
resource use drops significantly or entirely, leaving only the pro-
cedural resource to perform the task. This process explains why IF the goal buffer contains a dual task
multiple new tasks are hard to combine: not because an executive THEN add a goal to perform the choice task
process can attend to only one new task at a time, but because two and add a goal to perform the secondary task.
new tasks both heavily depend on declarative memory as a re-
source that they almost continuously need. Both goals are thus created and added to the set of active goals
maintained by threaded cognition; in other words, the goals are
both added to and maintained in the new multichunk goal buffer.
Core Assumptions for Threaded Cognition and In addition, a rule can indicate completion of the current goal by
Multitasking Performance simply removing the current goal from the set:
Building on the single-task assumptions above, threaded cogni- IF the goal buffer contains a choice task
tion adds several critical assumptions that describe the nature of and the motor response has been issued
THEN remove (terminate) the current goal (i.e., the choice-task goal).
multitasking performance. Although some of the theoretical as-
pects of these assumptions have been explored in other contexts
Each goal in the active goal set activates relevant rules and, in
(e.g., Kieras et al., 2000), threaded cognition provides a novel
executing these rules, produces a stream of processing across the
integration of these ideas in order to formulate a domain- system resources. We define a thread as all processing in service
independent, general multitasking model. Threaded cognition also of a particular goal, including procedural processing through the
extends the existing computational framework of ACT-R, resulting firing of rules and other resource processing initiated by these rule
in a new threaded version1 of the architecture that greatly facili- firings. For example, Figure 3a illustrates two threads for distinct,
tates the representation and modeling of multitasking behavior. In concurrent choice tasks: the visual–manual choice task seen earlier
fact, the threaded architecture allows for immediate integration of and an aural–manual choice task in which the stimulus is aural
two or more single-task models into a multitasking model and rather than visual. After a 50-ms delay in detecting a tone (details
immediate prediction of the resulting multitasking behavior.
Threaded Processing Assumption: Cognition maintains a set 1

The code that implements threaded cognition and all models described
of active goals that produce threads of goal-related processing in this article are publicly available at http://www.cs.drexel.edu/⬃salvucci/
across available resources. threads.
forthcoming in our later description of the dual-choice domain), combined with an aural–vocal thread, thus eliminating conflicts for
this thread fires a rule to encode the tone (i.e., determine whether perceptual and motor resources. However, the second thread still
the tone is high or low) and a second rule to generate the proper experiences a delay resulting from a conflict for the procedural
response. In this case, the aural–manual thread must wait for the resource, because the visual and aural encoding for the respective
manual resource to become free, thus causing a delay in overall threads complete at roughly the same time. Figure 3c illustrates the
reaction time for this task. Figure 3b illustrates a visual–manual same two tasks with less time needed for visual encoding; we see
here that no conflicts arise whatsoever, and thus perfect time
sharing is achieved. As another view of resource processing,
(a) Figure 4 provides the storyboard view of this final example. At this
point in our exposition, we have not refined our assumptions
enough to clearly define how the two threads would be executed
Visual-Manual Choice Task Aural-Manual Choice Task
on the resources. The subsequent assumptions further refine this
Procedural: Attend stimulus Aural: Detect tone specification to define exactly how multiple threads execute on the
Procedural: Attend tone processing resources.
Conceptually, a thread represents a particular stream of thought
Procedural: Respond to left stimulus Aural: Encode tone that is associated with some currently active task. A simple task,
such as one of the choice tasks illustrated above, likely has only
Manual: Press index finger one associated thread. However, many complex tasks could be
(waiting until manual module is free)
represented as a set of threads that collectively work toward a
Procedural: Respond to low stimulus higher level goal, but each individually focuses on a particular
aspect of that goal. Very complex dynamic tasks, such as driving
Manual: Press index finger or air-traffic control, are easily thought of as incorporating multi-
ple task threads. For instance, driving and air-traffic control in-
volve both monitoring of the environment for situation awareness
(b) and action in the environment to alter its state; although these
threads are clearly related in serving the high-level goal of safe
control (e.g., performed actions are often related to a monitored
Visual-Manual Choice Task Aural-Vocal Choice Task
event), they can also act independently and communicate infor-
Procedural: Attend stimulus Aural: Detect tone mation when needed. However, tasks need not be this complex to
Procedural: Attend tone
be amenable to a threaded approach. For example, models of list
memory typically incorporate task processes of both stimulus
Aural: Encode tone encoding and list rehearsal (e.g., Anderson, Bothell, Lebiere, &
Matessa, 1998) that could be viewed as distinct threads in service
Procedural: Respond to left stimulus (waiting until procedural is free) of the larger task goal. As another example, models of transcrip-
Procedural: Respond to low stimulus
tion typing (e.g., John, 1996) include, in essence, separate threads
Manual: Press index finger to read text and simultaneously type this text.
Vocal: Say "one" A threaded perspective on such tasks allows for a straightfor-
ward account of the kinds of flexibility exhibited in human mul-
titasking. In particular, independent threads are not inextricably
tied to one another in a processing sense but rather are combined
or separated however needed for maximum flexibility. For in-
(c) stance, in the transcription-typing example above, both the reading
Visual-Manual Choice Task Aural-Vocal Choice Task
and typing threads represent well-practiced skills that can easily
Procedural: Attend stimulus Aural: Detect tone operate independently of the other. In the driving or air-traffic
control example, it is sometimes the case that another person can
Procedural: Attend tone
assist with the monitoring task (e.g., a team of air-traffic control-
Procedural: Respond to left stimulus Aural: Encode tone
lers with assigned tasks, together managing a particularly busy
airspace, or a vehicle passenger who checks for oncoming traffic);
Manual: Press index finger Procedural: Respond to low stimulus
in this case, the original person can eliminate their own monitoring
thread and concentrate their efforts on performing necessary ac-
Vocal: Say "one"
tions. A threaded perspective of such tasks provides an account for
how people integrate their skills as necessary to adapt to the
demands of the current task environment.
Figure 3. Dual-choice task, expert behavior. a: Model timeline for
visual–manual and aural–manual choice tasks. b: Timeline with aural– Resource Seriality Assumption: All resources— cognitive,
vocal second task. c: Timeline with aural–vocal second task and faster perceptual, and motor— execute processing requests serially,
visual encoding. one request at a time.
Vis-Man Declarative Vis-Man Declarative

Goal Aur-Voc Goal Aur-Voc

Attend-
Attend Attend
stimulus
stimulus tone
(detected) (detected)
Aural Detecting Encode Visual Aural Encode Encoding Visual
tone stimulus tone stimulus
t = 0-50 ms t = 50-100 ms
The presence of a stimulus detected by the visual resource The detected tone and the Aural-Vocal goal activate the
combined with the Visual-Manual goal trigger the rule that rule that initiates encoding of the tone (i.e., determines
initiates visual encoding. Meanwhile, the aural module is the tone's pitch as needed for the choice response).
detecting the aural tone.
Vis-Man Declarative Declarative

Goal Aur-Voc Goal Aur-Voc

Attend- to
Respond Respond to
Press Say Pressing
leftstimulus
stimulus index finger low stimulus index finger
"one"
"O--" Low tone

Aural Encoding Visual Aural Visual
tone
t = 135-185 ms t = 220-270 ms
When visual encoding is complete, the pattern placed in the When aural encoding is complete, the low tone and
visual buffer and the Visual-Manual goal activate the rule the Aural-Vocal goal activate the rule that initiates
that initiates the finger press. The Visual-Manual goal has the vocal response "one" associated with the tone.
completed and terminates. The Aural-Vocal goal then terminates.
Figure 4. Storyboard view of the timeline with aural–vocal second task and faster visual encoding.
One of the most important core assumptions of threaded cogni- requested process at a time, forcing subsequent requests to wait
tion is that all resources operate sequentially, processing only one until completion of the current process. We prefer the latter
request (and thus serving only one thread) at a time. At first glance, exclusive-use approach for two reasons. First, the exclusive-use
this assumption may seem quite surprising, running counter to the approach represents a stronger, more limiting theory in that the
well-known degrees of parallelism in the brain’s neural circuitry. capacity approach can mimic the exclusive-use approach but not
We should make clear that although our characterization of human vice-versa; the capacity approach also requires additional param-
resources and their associated processes certainly represents an eters that define both the total capacity for a resource and the
approximation of the true system, we would argue that it provides capacity usage for any given process, a large space of new param-
a useful approximation in providing those distinctions most critical eters. We start with the stronger theory in order to evaluate its
to understanding multitasking at the level of abstraction addressed potential for accounting for behavior in common task domains.
in this article. In particular, our theory allows for the inclusion of Second, the exclusive-use approach generalizes in a straightfor-
parallelism at the level of multiple resources (e.g., parallel visual ward way across all cognitive, perceptual, and motor resources. In
and aural processing) and parallelism within each resource (e.g., contrast, in a parallel capacity-limited approach, the manner in
visual processing of scenes or parallel activation competition for which different resources utilize parallelism would depend on the
declarative retrieval) but requires sequential processing at the level particulars of that resource, notably in terms of whether the re-
of an individual resource—a separation which, for the types of source contains a serial bottleneck—for instance, any parallelism
tasks addressed in this article, nicely captures both the power and in the motor system must eventually be resolved to direct a single
limitations of multitasking behavior. physical resource (e.g., moving the finger to a location), and thus
To account for limitations on multitasking performance, theories such resources must resolve conflicts and serialize processes
have typically constrained resources in one of two approaches. somehow within the system. We thus prefer the unified exclusive-
One approach (e.g., Just et al., 1999; Wickens, 2002) places use approach and, for the sake of parsimony, require that all
capacity limits on resources such that each process occupies some resources abide by this approach (in contrast to, e.g., EPIC’s
percentage of the total capacity and that multiple processes can parallel cognitive processor but sequential motor system: Meyer &
execute on the resource as long as the total usage across processes Kieras, 1997a).
does not exceed the total capacity. A stricter approach (e.g., The most important implication of the resource seriality as-
Anderson et al., 2004) posits that resources can execute only one sumption arises in the procedural resource, where the assumption
forces the requirement that only one rule proceed at a given time. take significantly longer than the 50-ms rule-firing time for the
In general, as a production system matches and fires rules, situa- procedural resource.
tions may arise in which multiple rules could potentially fire at the The bottlenecks that arise from resource conflicts are often
same time (i.e., the conditions for multiple rules match the current associated with perceptual and motor resources, but they need not
state of the buffers). Some cognitive architectures such as EPIC be; in fact, one of the most interesting examples of resource
(Meyer & Kieras, 1997a) allow for multiple rules to fire simulta- contention arises in the declarative resource, particularly in cases
neously in these situations. In contrast, threaded cognition follows that involve practice and learning. Earlier we discussed the mech-
the spirit of the ACT-R cognitive architecture and requires that anism of production compilation, which transforms declarative
rules fire sequentially. Anderson et al. (2004) provided a neural instructions into procedural rules that perform a task. This mech-
argument for the single-rule requirement, namely that in cases in anism predicts that declarative memory retrievals are frequent in
which the current goal activates (i.e., matches) more than one rule, the early stages of learning but gradually drop out as the instruc-
inhibitory projections from the striatum to the pallidum allow a tions evolve into procedural rules. Generalizing this to the case of
single rule to dominate over the others and produce its selected multiple threads, threaded cognition may result in diminishing
action (see Graybiel & Kimura, 1995). In a sense, threaded cog- interference over time for two reasons: The learning thread grad-
nition generalizes ACT-R’s mechanisms to apply across a set of ually uses the declarative resource less frequently and avoids
active rules—that is, when multiple rules are activated by currently conflict with other threads that require this resource, and the
active goals, this same mechanism produces only a single domi- learning thread fires fewer production rules overall and thus re-
nant rule and only this dominant rule is allowed to proceed. quires less processing on the procedural resource.
Procedural seriality is critical to the workings of threaded cog-
nition because it posits the existence of a procedural bottleneck Resource Usage Assumption: Threads acquire and release
separate from potential bottlenecks in other resources. For exam- resources in a greedy, polite manner.
ple, returning to the dual-choice example illustrated in Figure 3b,
the visual–manual and aural–vocal choice tasks utilize different Further specification of how threaded cognition manages re-
perceptual and motor resources, and thus there is presumably no source processing requires some description of how the threads
potential for perceptual or motor bottlenecks in the dual-task case. themselves acquire and release resources. Using Kieras et al.’s
However, as the figure illustrates, there may still be a bottleneck or (2000) terminology, threaded cognition can be characterized as
conflict for the procedural resource. The procedural bottleneck is providing a “liberal” managerial style that allocates resources in a
essentially the central bottleneck that some empirical studies have “tolerant, laissez-faire” manner, in which threads largely manage
noted in dual-task performance (see, e.g., Pashler, 1994). Put in their own usage in requesting and freeing resources—as opposed
other terms, one could view procedural seriality as stating that to a “conservative” executive that allocates in a “strict regimented
cognition can process resource results and issue resource requests style.” Such a managerial style, however, could be abused by
for only one thread at a time; nevertheless, one thread’s resource processes that unfairly monopolize resources and thus creates a
processing can proceed in parallel with another as long as they do need for a “process etiquette” within the threads.
not both require procedural processing at the same time. Our formulation of threaded cognition defines such an etiquette
Although a serial procedural resource runs counter to some by requiring that thread usage of available resources be “greedy”
notable cognitive architectures, there is general agreement that and “polite.” Threads acquire resources in a greedy manner by
other system resources—perceptual, motor, and declarative re- requesting resources as soon as possible when needed. (Note that
sources— can be well characterized as operating in a sequential all resource requests occur in the actions of the rule currently
manner (though process execution may contain parallelism, such firing, and the firing of this action coincides with the acquisition of
as visual processing of an external scene). Our incorporation of the resource.) In contrast, a few related efforts have explored how
ACT-R’s perceptual and motor resources (Anderson et al., 2004; to model multitasking that involves more complex, nongreedy
see also Byrne & Anderson, 2001), which as mentioned largely solutions to resource contention. For example, Freed (1998) has
derive from those of EPIC (Meyer & Kieras, 1997a), follow the described a “sketchy planner” called APEX that has the ability to
assumptions of these architectures that these resources require explicitly delay, interrupt, and prioritize multiple task threads.
serial processing. ACT-R’s declarative resource, which handles Norman and Shallice’s (1986) supervisory attentional system al-
retrievals from declarative memory, is similarly assumed to require lows, in principle, for schemas that direct the same kinds of
serial processing. nongreedy planning and prioritization. Howes et al. (Howes, Vera,
The resource seriality assumption generates immediate predic- & Lewis, 2007; Howes, Vera, Lewis, & McCurdy, 2004) have
tions about resource conflicts in the case of multiple threads. described a “cognitive constraint modeling” methodology that
Returning to the example in Figure 3a, the two illustrated tasks enables exploration of strategic differences in multitasking behav-
share the manual resource as the necessary response modality. In ior through specification of behavioral constraints. We believe that
contrast to the short delay caused by the procedural conflict in these more complex multitasking behaviors may indeed be exhib-
Figure 3b, the manual resource conflict forces the aural–manual ited in some circumstances, and threaded cognition in principle
thread to wait until completion of the visual–manual thread’s does allow for complex multitasking through the presence of
motor response, creating a long delay in processing. This type of additional supervisory threads. However, we also believe that the
resource contention can, and typically does, produce more signif- greedy approach to resource contention is the more common one,
icant delays in the component tasks than contention for the pro- especially at the temporal grain size focused on here, as we
cedural resource, given that perceptual and motor processes can demonstrate in the later model simulations of a variety of tasks.
After acquiring a resource, a thread releases resources in a polite processing among threads: By ensuring that threads have a regular
manner by releasing, or freeing, a resource for other threads as opportunity to progress through the firing of procedural rules, the
soon as its processing is no longer required. As mentioned, for system allows all threads a chance to acquire resources and avoids
threaded cognition and its computational instantiation in ACT-R, starving any thread of processing time. When two threads exhibit
resources are in use when (a) they are currently performing a similar resource usage, the least recently processed policy results
processing request in service of some thread or (b) the results of a in an alternation of rule firings—the procedural resource can fire a
processing request (if any) remain in the resource’s buffer still rule for one thread while another thread’s peripheral processes
unused by the requesting thread. Rules that request a particular (vision, motor, etc.) are running, then vice-versa, and so on,
resource thus must check that the resource is not busy and that its achieving highly efficient parallelism between the two threads.
associated buffer is empty. After the thread has acquired the Our upcoming model simulations of dual-task behavior, including
resource, the requested process either terminates with no results those in which people exhibit perfect time sharing, demonstrate
(for a motor request) or places a result in the buffer (for a how such processing arises from threaded cognition. When two
perceptual or declarative request); in the latter case, a subsequent threads exhibit very different resource usage, such as when one
rule collects the result and thus clears the buffer. The resource, no task includes lengthy peripheral processes (e.g., motor move-
longer busy and with an empty buffer, can immediately accept new ments) whereas the other requires frequent procedural steps, the
requests from other threads. Thus, politeness is built into the least recently processed policy allows for the high-frequency task
system in that typical threads cannot monopolize resources be- to execute at high efficiency but still allows the low-frequency task
cause they require the results of the requested processes. (None- to acquire the procedural resource when attention is needed.
theless, a purposefully impolite thread could, in theory, request a
result and never use the result, leaving it in the buffer and main-
taining control of the resource; in practice, however, such behavior Relation to Other Theories of Multitasking
would not evolve naturally through the production compilation
Threaded cognition builds on a number of existing ideas and
process and could also be avoided through the use of an appropri-
theories and attempts to unify them under the auspices of a single
ate decay mechanism.)
computational theory. The earliest descriptions of single-channel
Figure 3 illustrates the greedy, polite threads that make up the
bottlenecks (e.g., Craik, 1948; Telford, 1931; Welford, 1952)
dual-choice task. In each of these figure panels, both threads
could be characterized within our theory if one considers the entire
greedily acquire resources when ready. For instance, in Figure 3b,
(short discrete) task, and all the resources involved (i.e., cognition
the visual–manual thread immediately encodes the stimulus and,
plus perceptual and motor resources), as a single processing chan-
upon completion, immediately issues the response; the aural–vocal
nel. Later theories that emphasize particular resources, such as
thread does not attempt to (nor can it) schedule or plan around the
bottlenecks that arise as a result of stimulus perception (e.g.,
other threads’ resource acquisitions but rather simply acquires
Broadbent, 1958) or response selection (e.g., Pashler, 1994), also
resources when available and waits when they are not. In Fig-
correspond nicely with aspects of the theory and its integration in
ure 3a, where the visual–manual and aural–manual threads contend
the cognitive architecture; for example, the visual perception and
for the manual resource, the completion of manual resource pro-
eye-movement aspects of the theory govern visual stimulus per-
cessing immediately frees the resource, allowing the aural–manual
ception, whereas the declarative-memory aspects of the theory,
rule that issues the motor response to proceed on the procedural
including recall and forgetting of memory chunks, govern at least
resource.
some components of response selection.
Conflict Resolution Assumption: When multiple threads con- The most closely related theories to our own are those that
tend for the procedural resource, the least recently processed attempt to integrate a number of modalities and resource con-
thread is allowed to proceed. straints into a unified framework. One well-known theory is the
multiple resource theory of Wickens (2002; see also Wickens,
Typically, when a thread desires a particular resource and that 1984), which characterizes dual-task interference in terms of four
resource is currently busy, that thread simply waits until comple- dimensions: stages (perceptual– cognitive vs. response), sensory
tion of the current process and then acquires that resource, as modalities (auditory vs. visual), codes (visual vs. spatial), and
illustrated in our earlier examples. In certain cases, however, when visual channels (focal vs. ambient). The theory incorporates a
that resource becomes free, there may be two or more threads methodology by which one can create a numerical conflict matrix
waiting to acquire the resource (possibly including the thread that that quantifies the resource conflicts between tasks across dimen-
just used and released the resource). We thus require some way to sions and allows for subsequent computation of the amount of total
specify, in these cases, which thread may proceed and acquire the interference (as a single numeric value). Although multiple re-
resource. source theory and its associated modeling methodology has the
Deciding which thread may acquire a resource reduces to the potential to capture multitasking behavior and interference at a
problem of deciding which thread may fire a rule on the procedural high level, it does not provide some of the many advantages of the
resource, because resource acquisition can occur only through a more detailed modeling possible in a cognitive architecture—for
rule firing. To make this decision, threaded cognition adopts a example, prediction of the fine-grained temporal sequence of be-
policy that the least recently processed thread—that is, the thread havioral events and prediction of actual quantitative measures to
that has least recently fired a rule on the procedural resource—is be compared with human behavior (e.g., measures ranging from
allowed to proceed. The primary motivation for the least recently reaction time to keystroke sequences to driver performance char-
processed policy is that it provides a parsimonious way to balance acteristics, as seen in the upcoming model simulations).
A related effort arises in Cooper and Shallice’s (2000) compu- allows multiple rule firings in parallel in contrast to our inclusion
tational implementation of contention scheduling (Norman & of a serial procedural resource. Nevertheless, our work follows
Shallice, 1986). Contention scheduling allows for triggering of very much in the spirit of Kieras et al.’s (2000) effort, and we have
appropriate actions through both top-down control and environ- examined and modeled some of the same tasks (described in the
mental stimuli—an approach very similar to our own—and this section on model simulations) to further compare and contrast our
action-selection mechanism can be the primary bottleneck in dual- approaches.
task performance. The computational model of contention sched- Threaded cognition arose from a unification of two recent
uling further fleshes out the original theory, with a demonstrated modeling approaches by the individual authors. Salvucci (2005)
application in the real-world domain of coffee preparation. At the explored an ACT-R general executive that incorporated queuing
same time, the model and its application focus on sequential action and timing mechanisms to achieve multitasking behavior. This
selection in this domain rather than multiple concurrent activities. approach led to more conservative multitasking because, unlike
In principle, this approach allows for multitasking if there are no threaded cognition, only one goal was active at a time, which led
resource conflicts between tasks, including both cognitive and to cases in which other goals could not proceed even though the
effector (perceptual and motor) resources. However, in the pres- currently active goal was waiting for a peripheral resource. This
ence of resource conflicts, the theory relies on higher level sche- work also required a new architectural extension for timing goals
mas in the supervisory attentional system to resolve the conflicts, on a queue, whereas threaded cognition requires no such exten-
and at this time there is no theory of the learning and behavior of sion. Taatgen (2005) modeled dual-task performance such that, in
such schemas. Thus, the contention scheduling model at present essence, both task goals were embedded into a single representa-
does not fulfill our desire of a general model of multitasking tion. Although this approach achieved more efficient multitasking
performance for arbitrary tasks. than that of Salvucci (2005), it relied on more tightly coupled task
A separate effort by Liu et al. (2005) explores the representation representations that made it more difficult both to learn tasks
of cognitive models as queuing networks within their QN-MHP individually and to perform tasks independently when desired. In
(Queueing Network-Model Human Processor) architecture (see addition, there were no constraints on when each task could
also Liu, 1996). The queuing networks provide a mathematical proceed (in contrast with threaded cognition’s preference for the
framework with which to express a parallel network of servers, least recently processed thread), and thus nothing prevented un-
each of which may process information serially or in parallel. As balanced processing or starvation of one task over another. We
such, the architecture provides a computational mechanism for view threaded cognition as incorporating the best of each ap-
concurrent task execution without the need for task-specific exec- proach—a straightforward, parsimonious theory that accounts for
utive processes. At the same time, the architecture employs “sep- the highly efficient yet highly flexible nature of multitasking
arate servers to handle distinct goal execution functions such as observed in human behavior.
goal prioritization, complex cognitive processing . . ., performance
monitoring, and procedure selection” (Liu et al., 2005, p. 49). In Key Claims and Predictions of Threaded Cognition
contrast, threaded cognition does not require specialized modules We can summarize the key claims of threaded cognition theory
for task priorities or procedure selection, relying solely on its as follows:
threading mechanism to execute concurrent tasks. Also, the queu-
ing networks leave open the question of serial versus parallel 1. Cognition can maintain and execute multiple active goals,
processing at the server level; threaded cognition is the narrower resulting in concurrent threads of resource processing.
and stronger theory in its commitment to resource seriality in
which resources process a single request at a time. 2. Threads can be characterized as alternating blocks of
The treatment of multitasking most closely related to our own is procedural processing (i.e., rule firings that collect infor-
that of Kieras et al. (2000), who studied production-system models mation and initiate new resource requests) and processing
of multitasking in the context of the EPIC cognitive architecture on peripheral resources (including perceptual, motor, and
(Meyer & Kieras, 1997a). In particular, they explored potential declarative memory resources).
characteristics of general executive processes that govern across
task domains, in contrast with customized executive processes that 3. Processing interference can arise on the central proce-
define domain-specific aspects of multitasking performance. In dural resource as well as on the declarative, perceptual,
some ways, one could characterize threaded cognition as, in their and motor resources.
terminology, a “liberal general executive with polite task pro-
4. Threads acquire resources greedily and release resources
cesses”: Task models are polite in the sense that they request
politely, which arises naturally from the characterization
resources only when needed and release resources when done,
of resources as modules and buffers.
allowing threaded cognition to be liberal in allowing task pro-
cesses to proceed whenever they can. However, threaded cognition 5. Cognition balances thread execution by favoring least
differs from the Kieras et al. conceptualization in that it does not recently processed threads on the procedural resource.
utilize rule-based executive processes to manage resources and
resolve resource conflicts but rather posits a parsimonious archi- 6. With practice, threads become less dependent on retrieval
tectural mechanism that performs this work for all tasks. In a of declarative instructions, reducing conflicts for both the
sense, threaded cognition introduces the function of a general declarative and procedural resources.
executive without the need for an explicit executive system or
process. Another important difference is that the EPIC architecture 7. Cognition requires no central or supervisory executive
processes; instead, multitasking emerges from the inter- These four domains were chosen specifically to illustrate the
action of autonomous process threads in conjunction with generality of the theory with respect to four conceptual axes, as
the key claims above. outlined in Table 1. First, the domains illustrate how concurrent
multitasking sometimes produces interference effects (i.e., de-
These claims give rise to several general, qualitative predictions graded performance) in one or both tasks, but sometimes it pro-
about the nature of dual-task performance. When two tasks require duces no such interference. In fact, two of the domains (dual
common perceptual or motor resources, dual-task performance for choice and driving) demonstrate how interference can either occur
one or both tasks will typically be impaired. Even when the tasks or not occur depending on specific task conditions. Second, the
utilize distinct perceptual or motor resources, dual-task perfor- domains illustrate how practice can change the resulting effects of
mance may be impaired as a result of potential procedural bottle- multitasking behavior over time. Third, the domains span both
necks. However, in the absence of procedural interference, dual- classic, well-studied laboratory tasks as well as realistic complex
task performance can be equivalent to single-task performance. In tasks, ensuring that the theory generalizes to real-world task do-
addition, with practice of either or both tasks, dual-task interfer- mains. Fourth, the domains illustrate how models can be reused for
ence will typically lessen as a result of the gradual compilation of predictive purposes such that a model of a particular task can be
declarative instructions into procedural form. These predictions integrated with new tasks to produce new predictions, an important
have all been supported by empirical work examining the central aspect of our integrated theory. We now describe each of the
bottleneck in dual choice tasks with distinct modalities (see modeling efforts in detail, which all together demonstrate how
Pashler, 1994), perfect time sharing in dual-task performance (e.g., threaded cognition can account for a broad range of multitasking
Hazeltine, Teague, & Ivry, 2002; Schumacher et al., 1999), and behavior.
practice and learning effects in dual-task behavior (Schumacher et
al., 2001).
At the same time, the computational realization of threaded Dual Choice: The Perceptual Refractory Period (PRP)
cognition provides much more than general qualitative predictions: and Perfect Time Sharing
It immediately predicts the behavioral process that results from
performing two or more tasks concurrently. One could begin with Dual-choice tasks represent the most widely studied task do-
two or more models of individual tasks, each independently de- main in the concurrent multitasking literature. Most dual-choice
veloped as a sound model of behavior for the respective domain. studies use the PRP paradigm, in which one task is given priority
Given such models, threaded cognition allows for simulations in over the other. Schumacher et al. (2001) have argued that people
which the model executes concurrently, generating predictions cannot fully exhibit their dual-tasking abilities in that paradigm
about multitasking performance. In addition, such models typically because of the priority between tasks. As a consequence, they
interact with a simulated environment that collects data and ana- argued, people may postpone the secondary task until the primary
lyzes behavior for many relevant observable measures, moving task is well underway, producing interference effects that can be
beyond simple reaction time measures. Thus, the space of testable attributed to the peculiarity of the instructions, instead of being due
predictions for threaded cognition is enormous: It claims that for to cognitive limitations. In order to test this idea, they conducted
any two or more independently validated task models, the com- dual-task experiments in which both tasks had equal priority. In
bined multitasking model as produced by threaded cognition will their Experiment 1, participants were given instructions for two
match the behavior of people performing these same tasks con- choice-reaction tasks. The first task was a visual–manual task, in
currently. In the next section, we sample from this space of testable which a circle would appear on the screen in one of three positions,
predictions and compare the theory’s predictions to empirical data and a finger corresponding to the position had to be pressed (left
for a representative set of task domains. position mapped to the index finger, middle position to middle
finger, right position to ring finger). The second task was an
aural–vocal task, in which a tone of low, intermediate, or high
Model Simulations of Representative Tasks
pitch was presented, to which participants had to respond “one,”
With the goal of testing our theory of threaded cognition, we “two,” or “three,” respectively. The tasks were first trained in a
chose four representative task domains in which to examine and single-task condition on Day 1 and were then tested in both
compare the theory’s predictions to human performance: dual single-task and dual-task conditions on Days 2–5. The results show
choice, tracking and choice, reading and dictation, and driving. that participants suffered from dual-task interference on the first
Table 1
Overview of Theoretical Coverage Provided by the Chosen Modeling Domains
Tracking and Reading and Driving and

Conceptual aspects of domain Dual choice choice dictation secondary tasks
Dual-task interference X X X
Lack of dual-task interference X X X
Practice effects X X
Laboratory tasks X X X
Real-world tasks X
Model transfer X X
days of the experiment but that this interference disappeared requires a second memory retrieval. The process timeline illus-
completely by Day 5. trates how the declarative representation of the task initially cre-
The model of this task is based on an earlier version reported by ates dual-task interference, primarily because declarative memory
Taatgen (2005), although that model used a specialized goal rep- becomes a heavily contended resource; the gray areas in Figure 5
resentation instead of the more general threading approach. At the indicate periods of time in which one thread must wait until the
start of the task the model initiates two threads, one for the other thread releases a resource, causing delays in task perfor-
visual–manual task and one for the aural–vocal task. We start mance. With practice, the production compilation mechanism de-
examining the model behavior when it is done learning, corre- scribed earlier gradually compiles the retrieval of declarative rep-
sponding to performance on Day 5, when people exhibited perfect resentations into task-specific production rules. This frees up
time sharing—when task performance is equally fast in single-task declarative memory as a resource, removes the bottleneck from the
and dual-task conditions. In fact, we have already used this model model, and eventually produces the interference-free multitasking
in our earlier examples—namely, the model depicted in the pro- shown in Figure 3c. Figure 6 shows the match between model and
cess timeline in Figure 3c and the storyboard view in Figure 4. The data (R2 ⫽ .96, root-mean-square error [RMSE] ⫽ .026).
only shared resource in this model is the procedural resource, but To support their claim that the PRP effect arises as a result of
this does not result in conflicts, because the two task threads need instructions, Schumacher et al. (2001) took the participants from
the procedural resource only at times when the other thread is Experiment 1 and put them in a PRP paradigm with the same tasks.
executing other (nonprocedural) processes. Therefore, after many The main difference was that they were instructed to give priority
trials of practice, there is no interference between the two tasks. to the aural–vocal task, instead of the equal priority of the original
There is interference, however, in the novice stage, as depicted task. In addition, the visual stimulus was presented 50, 150, 250,
in Figure 5. This figure shows the process timeline in the very or 500 ms after the aural stimulus (a delay period called the
initial stages of task performance, when each procedural step stimulus onset asynchrony or [SOA]). As expected, they indeed
requires time for one rule firing to retrieve the next task instruc- found the PRP effect—shown in Figure 7— even though the same
tion, time for the memory retrieval itself, and time for a second participants were perfectly capable of doing both tasks in parallel
rule firing that interprets and performs the task instruction. (We in Experiment 1. To model this variation of the task, we assumed
omit the storyboard view of this model because of space con- that instead of initiating the two threads at the beginning of each
straints; however, one can imagine extrapolating the storyboard in trial, only the thread of the primary task (aural–vocal) is initiated
Figure 4 into extra rule firings and declarative retrievals as illus- and that the thread of the secondary task is initiated only when the
trated in Figure 5.) The assumption is that people do not have procedural portion of the primary task has completed—that is, at
task-specific rules when they enter the experiment but rather the onset of a vocal response. We otherwise assume a model that
memorize instructions (as done in the experiment) in declarative has been trained on the task in Experiment 1, to mimic the same
form, then interpret these instructions as needed during task exe- situation as that experienced by the experiment participants. Fig-
cution. Another assumption is that people can immediately map ure 7 shows that this slight modification of the model indeed fits
screen positions onto fingers but that mapping tones onto numbers the data from Schumacher et al. (R2 ⫽ .95, RMSE ⫽ .016).
In Experiment 1, visual processing was much faster than
auditory processing, ensuring that the moments of response
Visual-Manual Choice Task Aural-Vocal Choice Task selection never coincided. In contrast, Schumacher et al.’s
(2001) Experiment 3 increased the difficulty of the visual–
Procedural: Retrieve instruction Aural: Detect tone
manual task to see whether simultaneous response selection for
Declarative: Retrieve instruction both tasks would lead to interference. Instead of having a
(waiting until declarative is free)
Procedural: Attend stimulus congruent mapping from visual positions to finger locations,
Procedural: Retrieve instruction
participants had to reverse the order: Now, the leftmost visual
stimulus mapped onto the rightmost finger, and so on for all
Declarative: Retrieve instruction
fingers. The results, shown in Figure 8, show that this manip-
(waiting until declarative is free)
Procedural: Attend tone ulation indeed makes the task more difficult. Moreover, even at
Procedural: Retrieve instruction Day 6, people do not completely succeed in perfect time shar-
Aural: Encode tone ing. A more detailed analysis by Schumacher et al. showed that
Declarative: Retrieve instruction
some individuals achieve perfect time sharing by Day 6, but
Procedural: Respond to left stimulus (waiting until declarative is free) others do not. To model these data, we took the model of
Procedural: Retrieve instruction Experiment 1 and changed the instructions for the visual–
Manual: Press index finger
Declarative: Retrieve instruction manual task: Because the stimulus–response mapping is now
incongruent, we required the model to retrieve stimulus–
Procedural: Retrieve tone response
response mappings from declarative memory, reflecting the
Declarative: Retrieve low = "One" assumption that an incongruent mapping requires conscious
Procedural: Respond to low stimulus effort to resolve. The model fits are included in Figure 8 (R2 ⫽
.98, RMSE ⫽ .020).
Vocal: Say "one" Overall, the Schumacher et al. (2001) dual-task paradigm pro-
vides an example in which the contended resources include de-
clarative memory and, in Experiment 3, cognitive processing in
Figure 5. Dual-choice task, novice behavior: Model timeline. terms of rule firings. In this case, the effect of declarative memory
Experiment 1
0.8
Dual Aural−Vocal (data)
Dual Aural−Vocal (model)
Single Aural−Vocal (data)
Single Aural−Vocal (model)
Dual Visual−Manual (data)
Dual Visual−Manual (model)
Single Visual−Manual (data)
0.6 Single Visual−Manual (model)
Reaction time (s)
0.4
0.2
0.0
2 3 4 5
Day
Figure 6. Dual-choice study: Experiment 1 data and model results.
retrievals can be initially quite substantial, but it also gradually bottleneck. Their conclusion was that given enough training, par-
drops out with practice. The effect of cognitive processing is ticipants can achieve almost perfect time sharing but that there still
relatively small in comparison, because the opportunity for a remains a small residual interference in the order of 10 ms. These
resource overlap is very small if the response selection step takes effects have subsequently been modeled by Anderson, Taatgen,
only 50 ms. Hazeltine et al. (2002) have conducted a further series and Byrne (2005) using models very similar to the ones discussed
of experiments that have explored this central response selection here, again with the exception that they used a control structure
Experiment 2 (PRP)
0.7
Aural−Vocal (data)
0.6
Visual−Manual (data)
Aural−Vocal (model)
Visual−Manual (model)
0.5
Response time (s)
0.4
0.3
0.2
0.1
0.0
0.0 0.1 0.2 0.3 0.4 0.5
SOA (s)
Figure 7. Dual-choice study: Experiment 2 data and model results. PRP ⫽ perceptual refractory period;
SOA ⫽ stimulus onset asynchrony.
Experiment 3: Visual−Manual Experiment 3: Aural−Vocal
1.2
1.2
1.0
1.0
Reaction time (s)
Reaction time (s)

0.8
0.8
0.6
0.6
0.4
0.4
Dual task (data) Dual task (data)
0.2
0.2
Dual task (model) Dual task (model)
Single task (data) Single task (data)
Single task (model) Single task (model)
0.0
0.0
2 3 4 5 6 2 3 4 5 6
Day Day
Figure 8. Dual-choice study: Experiment 3 data and model results.
that was specifically designed to accommodate the two tasks. recall resulting from the concurrent task and concluded that
These models are also quite amenable to being cast in terms of our episodic recall is part of the central bottleneck. However, an
proposed threaded approach. account with dual bottlenecks of procedural and declarative can
The dual-task models show that although the procedural give a better explanation. The difference between the Rohrer
resource is the central resource in our theory, declarative mem- and Pashler experiment and earlier studies was that their
ory can also serve as an important source of interference. choice-reaction task had arbitrary mappings, making it neces-
Several related studies have explored the role of declarative sary to retrieve these mappings from memory—therefore both
memory in multitasking behavior. Johnston, McCann, and Rem- tasks required declarative memory access, producing interfer-
ington (1995) have demonstrated that there can be two bottle- ence. If the choice-reaction task does not involve memory
necks in a choice-reaction task, which they call input attention, retrieval, as in the earlier studies, threaded cognition predicts no
associated with initiating visual perception, and central atten- interference on free recall.
tion, associated with response selection. In our view, however,
these are both associated with the procedural resource, because Tracking and Choice: Dual-Task Interference in a
a choice-reaction task requires two production rules to make a Continuous Task
response. Pashler and Johnston (1998) did not rule out that there
is more than one bottleneck but maintained that a single bot- The dual-choice task domain involves two discrete, one-shot
tleneck is the most parsimonious account to fit the data. How- tasks that last less than a second and thus provide an interesting
ever, they did not consider the role of declarative knowledge as but limited window into multitasking behavior. By replacing
a source of task knowledge to explain novice behavior but one of the tasks with a continuous task that requires constant
instead considered retrieval from long-term memory to be part monitoring and response, we can explore the effects of multi-
of one bottleneck. The basis for this is a study by Carrier and tasking in the context of an interaction between a discrete and
Pashler (1995) that investigated whether memory retrieval is an continuous task. One domain that has received some attention in
independent bottleneck. In a dual-task choice paradigm, they recent literature (e.g., Chong, 1998; Kieras et al., 2000; Lalle-
found that memory retrieval in one task is postponed by re- ment & John, 1998) is that of the tracking and choice task: A
sponse selection in another task and concluded that memory person tracks a moving cursor and attempts to keep it centered
retrieval is part of a unitary bottleneck. However, in their on a target, and occasionally a choice stimulus appears that
conceptualization of the task, perception itself initiates the requires a choice response. Like past efforts, we base our
memory retrieval, whereas in our framework memory retrieval simulations on the empirical results of Martin-Emerson and
is always initiated by a production rule. The production rule that Wickens (1992), who ran a version of this tracking and choice
initiates memory retrieval can fire only after the response- task. In their experiment, participants tracked the moving target
selection production rules have finished, explaining the inter- in either an easy or a hard condition; the difficulty of tracking
ference results of Carrier and Pashler (1995). related to the pseudo-random function that perturbed the tar-
In another study, Rohrer and Pashler (2003) let participants get’s position. In addition, occasionally an arrow would appear
do a free-recall task and a choice-reaction task in parallel. on the screen and the participant would respond with a keypress
Contrary to earlier studies (e.g., Baddeley, Lewis, Eldridge, & as to whether the arrow pointed left or right. The position of the
Thomson, 1984), they found a substantial degradation in free arrow was manipulated in the experiment such that it appeared
at differing degrees of visual angle down from the tracking target, and complete the update. When the choice thread sees the
target—stimulus offset—in order to test the difficulty of the arrow stimulus (the third rule firing), it visually encodes the arrow
choice task as dependent on its distance from the continuous and generates the appropriate key response. During this time, the
tracking task. choice thread occupies the visual and then the manual resources,
We modeled the tracking task with three simple rules that delaying the tracking movement (which requires the manual re-
find the moving cursor, move the cursor to the target, and source for movement and the visual resource to guide the move-
reiterate this process. We modeled the choice task with a ment). This delay creates interference in the tracking thread, im-
similarly simple model: one rule that attempts to find the arrow, pairing tracking performance for a short time and causing a larger
another rule that recognizes when it was not found and iterates, deviation of the cursor from the target.
one rule that attends the arrow when found, and another two Figure 10 shows two measures of performance: (a) tracking
rules that generate the response for a left or right arrow. The error as a function of stimulus offset and (b) choice response
model runs both component task models as individual threads, time as a function of stimulus offset. In Figure 10a, we see a
and because they vie for visual attention, this resource is being large effect of tracking difficulty in the empirical data (solid
shared between the threads. The model also shares the manual lines), as well as a slight but steady increase in error as the
resource in that it uses its right hand for movement and its left offset between tracking target and choice stimulus increases.
hand for choice response (note that the cognitive architecture The model results (dashed lines) show the same effects and
currently specifies a single manual module, with both hands as quantitatively match the data well (R2 ⫽ .97, RMSE ⫽ .68): It
a single shared resource). The task environment used by the exhibits the difficulty effect because at a constant rate of
model was based on that implemented by Chong (1998) as a updating the tracking task, the cursor in the hard condition is
duplicate of the original experiment; most important, the per- perturbed by a greater distance, and it exhibits the stimulus
turbation functions for the easy and hard conditions were copied offset effect because of the increased time needed to encode a
exactly to ensure that quantitative comparison of the resulting visual object (the choice stimulus) at a farther distance (a
measures was made on the same scale. Results were compiled prediction of ACT-R’s eye-movement model: Salvucci, 2001a).
from simulation runs with 16 trials for each stimulus offset and In Figure 10b, the empirical data show an effect of stimulus
for each tracking difficulty condition, with no architectural offset on choice reaction time but no effect of tracking diffi-
parameters estimated to produce model fits. culty. The model results (R2 ⫽ .74, RMSE ⫽ .06) duplicate this
To illustrate the workings of the model, Figure 9 shows a effect of stimulus offset for the same reasons as for tracking
process timeline representing resource processing immediately error but also do not exhibit difficulty effects because difficulty
before and after the presentation of the arrow stimulus. The first does not affect encoding of the choice stimulus.
block of the tracking thread shows how tracking comprises an These results compare well with results obtained by other
iteration of three rules that find the cursor, move it back to the researchers working with different modeling frameworks, such
as Kieras et al. (2000) with the EPIC cognitive architecture and
Chong (1998) with a hybrid EPIC–Soar architecture. Lallement
Tracking Choice and John (1998) performed an interesting comparison of these
models and one of their own, showing that all the models, and
Procedural: Look for arrow
(waiting until visual location is free) the “modeling idioms” on which they are based, produce be-
Procedural: Arrow not found havior that corresponds quite well to the empirical data. Our
Procedural: Find cursor results are comparable to the results of these other models, but
Procedural: Move cursor
our approach has one major advantage over previous models:
Whereas the others required explicit specification of when to
Manual: Visual:
Move cursor Watch cursor
(waiting until visual location is free) switch between tasks, resulting in fairly complex behavioral
flow diagrams (see, e.g., Kieras et al., 2000), our model is a
straightforward integration of two simple component models
Procedural: Done iteration with no explicit specification of task switching, instead allow-
Procedural: Look for arrow
ing threaded cognition to manage and execute the two tasks.
(waiting until visual location is free)
Procedural: Encode arrow
Reading and Dictation: Perfect Time Sharing in
Procedural: Find cursor Visual: Encode arrow
Continuous Tasks
The tracking and choice task above addresses a case in which a
Procedural: Respond to arrow
(waiting until visual continuous task (tracking) exhibits interference when performed
and motor are free) Manual: Press key
concurrently with another task (choice). For some domains, how-
ever, studies have shown that people can achieve perfect time
Procedural: Move cursor sharing for continuous tasks just as they sometimes achieve it for
simpler tasks, as in the dual-choice task domain seen earlier. The
Manual: Visual:
Move cursor Watch cursor classical case is a study of reading and dictation run by Spelke,
Hirst, and Neisser (1976). In this experiment, 2 participants went
through an 85-day experiment in which they simultaneously had to
Figure 9. Tracking-choice study: Model timeline. read stories for comprehension and write down dictated words.
25 between 700 and 5,000 words and were selected from Ameri-
can, English, and translated European writers. On each daily
session, participants had to read three stories. While partici-
20
pants read the stories, the experimenter read words drawn from
Kucera and Francis (1967), which participants had to write
Tracking Error (pixels)
15
down without looking at their handwriting. After a participant
completed writing a word, the experimenter read the next word.
Once a participant finished reading a story, they were tested for
10 either comprehension of the story (10 times/week) or recogni-
tion of the dictated words (4 times/week). In addition, once per
Human Easy
Human Hard
week participants received a control trial in which they did only
5 Model Easy the reading task. Accuracy on both story comprehension and
Model Hard
recognition did not change much with practice, nor did the
writing speed, which was constant at about 10 words/minute.
0
0 3 6 10 13 16 19 22 26 29 32 35 The main change in performance was reading speed, as ana-
Stimulus Offset (°) lyzed for the study’s 2 participants, John and Diane. John,
whose normal reading speed was 480 words/minute, slowed
1.2
down to 314 words/minute in the 1st week but improved to 451
words/minute in the 6th week. Diane’s normal reading speed
1.0 was 358 words/minute. Her speed in Week 1 was 254 words/
minute, but this improved to 333 words/minute in Week 6.
To model the reading aspect of the task, we used an existing
Response Time (s)
.8
ACT-R model of reading developed by Lewis and Vasishth
(2005). Lewis and Vasishth’s model builds up a syntactic parse
.6
tree of a sentence by attaching new words into this tree as they
are read. It has a fairly straightforward pattern of behavior: For
.4
Human Easy each word to be read, the model visually encodes the word,
Human Hard retrieves its lexical item from declarative memory, retrieves a
Model Easy
.2 syntactical node in the parse tree from declarative memory to
Model Hard
which the word has to be attached, creates a new syntactical
.0 node with the word just read, and attaches it to the retrieved
0 3 6 10 13 16 19 22 26 29 32 35 syntactical node. Because the visual module can read the next
Stimulus Offset (°) word while the current word is processed, reading and parsing
a word takes slightly over 200 ms using Lewis and Vasishth’s
Figure 10. Tracking-choice study: Data and model results for tracking
error (a) and response time (b).
parameter settings. The model we constructed for taking dicta-
tion is similarly straightforward: it aurally encodes the spoken
word, retrieves the spelling of the word from declarative mem-
ory, and then writes the word one letter at a time. Although
Initially, reading speed was affected by the dictation task, but with writing is a highly trained skill, writing without visual guidance
practice, reading speed approached normal levels, as measured by is not; the model therefore requires a declarative retrieval for
average reading speed without dictation. The dictation speed was
each letter, where the retrieval is assumed to contain the move-
not affected by practice.
ment pattern needed to write down the letter. With practice,
The Spelke et al. (1976) study is interesting because it takes
production compilation produces rules for each letter that pro-
two tasks in which participants are already skilled and combines
duce the movement pattern right away. The only estimated
them. The disadvantage of the study is that the data are noisy
parameter in the dictation model was the manual time to write
(the participants differed considerably in reading speed) and the
reporting does not describe detailed behavior on a trial-by-trial a letter, which was set such that writing speed averaged 10
basis. More modern studies, such as those by Schumacher et al. words/minute.
(2001) and Hazeltine et al. (2002) addressed earlier, provide Figure 11 shows the operation of the model once it has reached
better insight into the exact timing of dual-task perfect time the expert stage. The reading model goes through the cycle of steps
sharing through the use of simpler choice tasks. At the same indicated above, and the dictation model listens to words and
time, such experiments are less representative for real-life mul- writes out the letters one at a time. The gray areas indicate that
titasking, because the choice tasks are artificial and trained in there is some interference each time a new word is dictated, but
the context of the experiment. Thus, the Spelke et al. study once the word is written out (which takes, on average, 6 s), there
provides a complementary view of more realistic multitasking, is little interference. The novice model, however, needs two pro-
albeit with less detailed timing data. duction rules and a declarative retrieval in the dictation task for
In the Spelke et al. (1976) experiment, participants had to every written letter (instead of the “initiate-movement” produc-
read stories while taking dictation. Stories varied in length tion), creating substantially more interference. The result is that the
Reading Dictation
Procedural: Attend word
Visual: Encode word

Aural: Detect word
Procedural: Do lexical access
Procedural:
Retrieve syntax Vision:
Encode word
Declarative:
Syntax element
(waiting until procedural is free) Procedural: Attend word
Procedural: Attach syntax Aural: Encode word
Procedural: Retrieve word

(waiting until procedural is free)
Procedural: Found spelling
Procedural: Do lexical access (waiting until procedural is free)
(waiting) Vision: Procedural: Initiate movement

Procedural:
Encode word
Retrieve syntax
Declarative:
Syntax element
Procedural: Attach syntax

Procedural:
Retrieve syntax
Vision:
Encode word
Declarative:
Syntax element
Manual: Write letter

Procedural:
Encode word
Declarative:
Syntax element

Procedural:
Encode word
Declarative:
Syntax element
Procedural: Initiate movement

Manual: Write letter
= Declarative: Lexical access (reading) or retrieve spelling (dictation)
Figure 11. Reading-dictation study: Model timeline.
model’s single-task reading performance is 298 words/minute. In Driving: Transfer and Integration in a Real-World
the dual-task situation this performance slows down to 229 words/ Domain
minute in Week 1 but then speeds up to 270 words/minute by
Week 6. Figure 12 shows a comparison between the model and the The three domains modeled to this point all address behavior in
2 participants for each of the first 6 weeks of the experiment, with the laboratory with the goal of elucidating the finer timing aspects
performance scaled to maximum reading speed. Threaded cogni- of multitasking behavior; even the reading and dictation task,
tion nicely captures the adaptation over time in the course of the which incorporates two common real-world skills, combines these
reading and dictation study. skills into a somewhat unnatural, certainly uncommon task do-
mated three parameter values for the current simulation (see Sal-
1.1
Reading speed as proportion of single task
1.0 vucci, 2005, for parameter descriptions): minor adjustments to the

steering scaling factor (0.85) and the stability scaling factor (3.0)
to account for slight differences in simulation environments and a
0.9
change to the accelerator-to-brake movement delay (600 ms) to

represent the minimum movement time (Lee, McGehee, Brown, &
0.8
Reyes, 2002) plus 200-ms motor preparation. The final modifica-

tion to the model relates to its timing aspects during stable control.
0.7
The original model incorporated a delay such that in stable situa-

tions—that is, when the car is near the lane center with low lateral
0.6
Model velocity—the model delayed the next control update for 500 ms;
Diane
John this delay represents a satisficing assumption in the original model
0.5
that in stable situations, continual control is not required, and this

1 2 3 4 5 6 assumption was found to nicely capture behavior. The driver
model used here incorporates the same assumption. However, two
Week
of the task domains below (driving and choice and driving and the
Figure 12. Reading-dictation study: Data and model results. sentence-span task) used an additional task that required the driver
to brake upon onset of a visual stimulus on a lead vehicle; because
of this additional task, we assumed a 50% probability of breaking
main. To extend our representative set of tasks to include a more out of the delay period at each potential rule firing, thus represent-
complex real-world task, we now examine driving in combination ing the fact that the human drivers had reason to pay further
with several possible secondary tasks. One major benefit of the attention to the driving task beyond those needed for basic control.
driving domain is that it allows us to apply the theory to a truly In other work, we have explored how appropriate delay times can
real-world task that many people experience on a daily basis. be learned and can adapt over time (Salvucci, Taatgen, & Kush-
Another benefit is that it allows us to demonstrate the potential for leyeva, 2006), but for simplicity we do not include these mecha-
transferring a single model of behavior across multitasking do- nisms in the current modeling effort.
mains, integrating this model with models of other tasks to predict Driving and choice. Recently Levy, Pashler, and Boer (2006)
the behavior that results in a dual-task condition. ran a study that combines a basic choice task with the very
Driving alone. The starting point for our treatment of driving common real-world task of driving—in essence, a PRP task in the
is a computational model of driver behavior that has been tested in context of a real-world driving task. Specifically, their task in-
a separate context. The ACT-R integrated driver model (Salvucci, volved following a lead vehicle while occasionally responding to
2005, 2006) can navigate a typical multilane highway environment
and has been compared against human driver behavior with respect
to performance measures for subtasks such as situation monitoring, Driving Choice
curve negotiation, and lane changing. In this article, all tasks
involve only the control (steering and acceleration– braking) as- Procedural: Find road near point
pects of the driver model, not the lane changing aspects, and thus Procedural: Find road far point
we import the core elements of the driver model that implement Procedural: Steer & accelerate
control. The driver model in essence boils down to four rules that
Procedural: Note vehicle is stable Aural: Detect tone
iterate in sequence, each iteration producing one “update” in which
(waiting until procedural is free) Procedural: Wait for next tone
steering angle and acceleration– deceleration pedals are potentially
adjusted. The first two rules find the “near point” and the “far Procedural: Find road near point
(waiting 100 ms)
point” of the current lane (see Salvucci & Gray, 2004); these two Procedural: Find road far point
points provide information about the nearby and upcoming lane (waiting until procedural is free) Procedural: Detect next tone
configurations (respectively), subsequently used by the model in Procedural: Steer & accelerate Aural: Detect tone
its calculations of steering angles. The third rule sends the motor 1
(waiting until procedural is free) Procedural: Respond to stimulus
commands to specialized motor modules for steering and pedal
movement and also directs visual attention to encode the object at Procedural: Note vehicle is stable Manual: Press key
the far point (i.e., a lead vehicle or a road point). The fourth rule Procedural: Find road near point
checks for the stability of the vehicle: If the vehicle is not suffi- Procedural: Find road far point
ciently stable as defined by its current lateral position and velocity, Procedural: Steer & accelerate
the process iterates immediately; if the vehicle is sufficiently 2
stable, the process iterates after some delay (defined shortly). = Manual: Steer; Pedal: Acc/Decelerate;
Visual: Encode far point object
These four rules are illustrated in the process timeline in Figure 13.
= Lead-car brake lights appear (SOA = 150 ms)
In addition, the past versions of the driver model were devel- = Brake lights noted during visual encoding
1
oped in earlier versions of the ACT-R architecture, and thus we 2 = Model overrides accelerator, produces brake response
imported the ACT-R 5.0 model from Salvucci (2005) and made
some implementation-related, nontheoretical changes to the model Figure 13. Driving-choice study: Model timeline. SOA ⫽ stimulus onset
to work with the newest architecture, ACT-R 6.0. We also reesti- asynchrony.
a choice task; the choice task had four conditions, with either a 13), the normal acceleration process is overridden (as per task
visual or auditory choice stimulus and either a manual or vocal instructions) and the braking process begins instead. The braking
response. In addition, a specified time after the stimulus onset— motor movement completes after a delay of 600 ms following
namely the SOA time described in the dual-choice section—the motor initiation; meanwhile the choice thread has terminated, and
lead vehicle’s brake lights would occasionally light up and require normal driving resumes. Note that after the onset of the visual
that the driver press the brake pedal. Thus, the task in essence braking stimulus (indicated by the star in Figure 13), the driving
involved three tasks, namely the discrete choice task, the discrete thread waits twice for the procedural resource, causing a 100-ms
braking task, and the continuous driving (steering) task, lending delay in this thread, and this delay is the source of the PRP effect
the task more external validity as a realistic task and providing an noted by Levy et al. (2006) in their study.
excellent transition to our subsequent simulations involving the Figure 14a shows the human and model reaction times (R2 ⫽
driving domain. .97, RMSE ⫽ .04) for the choice task and the braking task as a
Given the driver model, we adapted it to the Levy et al. (2006) function of SOA, analogous to our earlier analysis of the PRP
task as follows. First, we extended the model such that if brake effect in the dual-choice task. As in the earlier analysis, reaction
lights appear on the vehicle, it bypasses its default control of the times for the first task, the choice task, are unaffected by SOA.
accelerator and taps on the brake pedal instead. Second, we de- However, reaction times for the second task, the braking task, are
veloped a simple model for the choice task, following the same larger for short SOAs: For both human and model data, we see the
rule pattern as in the previous choice domains: When the stimulus largest increase for an SOA of 0 ms (simultaneous onset), a smaller
is detected, the model encodes the stimulus (either visual or increase for an SOA of 150 ms, but no increase at SOAs of 350 and
auditory) and presses a button that corresponds to the stimulus. 1,200 ms. Clearly the cognitive processing required early in the
However, Levy et al. used a slight variant to typical PRP studies in braking task conflicts with that of the choice task for short SOAs
that the stimulus was actually a single or rapidly repeated double and thus must be delayed for a short time, whereas this conflict
instance of a visual or auditory stimulus: One choice was presented does not occur for larger SOAs.
as a single visual flash on the lead vehicle’s rear-view window or As another view of these data, Figure 14b shows the reaction
a single auditory tone, whereas the second choice was presented as times for the four combinations of input modalities (auditory and
two rapidly repeated flashes or tones. In addition, for the manual visual) and output modalities (manual and vocal), for both the
response, the driver was required to press the input button once for human drivers in the experiment and the model (R2 ⫽ .95,
a single instance of the stimulus and twice for a double instance of RMSE ⫽ .04). The visual conditions require more time than the
the stimulus. The model for this task follows this procedure: After auditory conditions because of contention for the visual modality
noting the stimulus, it waits 100 ms (the interstimulus interval used with the driving task, and the vocal conditions require more time
in the original experiment) and then notes whether there is a because of the delays of speaking the responses and the subsequent
repetition; for the manual response to two instances of the stimu- recognition. However, the most interesting aspect of these results
lus, the model simply presses the button twice. For the vocal is the effect of multitasking. The first two blocks of results show
response to either stimulus, the model simply assumes a spoken the choice reaction time in the single-task (no braking) and dual-
one-syllable utterance to represent the responses “one” and “two.” task (while braking) conditions; for both humans and the model,
Finally, because the original experiment used a voice recognition the additional braking task incurs no additional reaction time,
system to collect vocal responses, we estimated a parameter to essentially mirroring the results for Task 1 in the dual-choice
represent the delay in the recognition system, estimated at 300 ms. domain. In the third block of results, we also see that input and
The final model used the driver model as one thread and the choice output modality had no effect on the braking time in the multi-
model as a second thread. Results were collected from simulations tasking conditions—in other words, braking performance while
with eight trials for each condition (auditory–manual, auditory– multitasking, as indicated in the PRP curves in Figure 14a, was not
vocal, visual–manual, visual–vocal) and for each SOA (0, 0.15, affected by input and output modalities.
0.35, and 1.20 s). Driving and phone dialing. By combining driving and choice
As before, we can examine a detailed process timeline to illus- in the previous model, we took one step toward examining a
trate how the two tasks are executed together, shown in Figure 13. realistic, nonlaboratory task domain. In this model, we make both
As mentioned, the driving process involves a repeated firing of tasks as realistic as possible to address a common, and unfortu-
four rules that find the near and far points, perform the steering and nately dangerous, real-world scenario: dialing a phone while driv-
acceleration (or deceleration) actions, and check for stability on the ing. Numerous studies have examined the effects of phone dialing
road. The figure illustrates driving and the choice task performed while driving (e.g., Alm & Nilsson, 1994; Brookhuis, De Vries, &
concurrently, with an assumed SOA of 150 ms. First, for the De Waard, 1991; McKnight & McKnight, 1993; Salvucci &
choice thread, the audition module detects a tone and waits for the Macuga, 2002), and there is general agreement that dialing as a
stimulus interval to hear the next tone. Once the presence or secondary task has significant potential to result in driver distrac-
absence of a second tone is noted, the choice response is made. tion with respect to the primary task—that is, decreased perfor-
Meanwhile, the visual stimulus of the lead vehicle’s brake light mance in the driving task. This domain also benefits our theory in
appears after the SOA interval (indicated by the star in Figure 13); that it provides a first test of threaded cognition for two nondis-
the subsequent visual encoding that occurs as part of the driving crete tasks: the continuous driving task and a dialing task that
thread encodes the presence of this brake light. The driving thread requires several seconds to complete and thus cannot be completed
then continues as usual during the choice response motor execu- as a one-shot task, as could the choice tasks.
tion, but when the acceleration adjustments are made for this We focus our modeling efforts on a recent study of driver
second driving update (indicated by the arrow labeled 2 in Figure distraction (Salvucci, 2001b) performed in a realistic driving sim-
Human Model
1.2
1.0
Reaction Time (s)
.8
.6
Brake Single Task Brake Single Task
.4 Brake Dual Task Brake Dual Task

Choice Single Task Choice Single Task
.2 Choice Dual Task Choice Dual Task
.0
0 .15 .35 1.20 0 .15 .35 1.20
SOA (s) SOA (s)
Human Model
1.2
Auditory-Manual Visual-Manual Auditory-Manual Visual-Manual
Auditory-Vocal Visual-Vocal Auditory-Vocal Visual-Vocal
1.0
Reaction Time (s)
.8
.6
.4
.2
.0
Choice Choice Brake Choice Choice Brake
Single Task Dual Task Dual Task Single Task Dual Task Dual Task
Figure 14. Driving-choice study: Data and model results for reaction time by SOA (a) and reaction time by
input– output modality (b). SOA ⫽ stimulus onset asynchrony.
ulator with a front half of a Nissan 240sx interfaced with a for each condition and updated them to the newest version of
simulated highway environment. In the study, drivers had to steer ACT-R and at the same time removed any code that specified the
down an open road while their velocity was held constant, and they explicit switching, allowing threading to handle all switching
occasionally dialed a seven-digit phone number in four conditions: between tasks. The only parameter assumptions made for the
full–manual, in which they pressed each digit; speed–manual, in dialing models were that all spoken digits were one syllable and all
which they pressed a single “speed number” associated with the spoken phrases for the speed–voice condition were three syllables.
phone number; full–voice, in which they spoke all seven digits and To begin dialing while driving, the full model simply created the
heard them repeated for confirmation; and speed–voice, in which particular dialing control state that corresponded to the desired
they spoke a single phrase (e.g., “office”) associated with the dialing method, thus creating a new dialing thread that executed
number and heard it repeated for confirmation. Before dialing in along with the driving thread. Results were compiled from three
all conditions, drivers were required to press a special power simulation runs, each with 16 dialing trials performed without
button to activate the phone, and after the manual conditions, driving and 16 trials performed while driving.
drivers were required to press the send key to complete dialing. Figure 15 shows a process timeline for dialing a seven-digit
The detailed results appear below, but in short, the full–manual number as two blocks of three and four digits (following the North
condition produced the most distraction, the speed–manual condi- American convention of a three to four chunking of the phone
tion produced less but significant distraction, and both voice con- number). For a new block of numbers, the dialing thread first
ditions produced no significant distraction. retrieves a declarative chunk that represents the block, then re-
The previous model of behavior for this task (Salvucci, 2001b) trieves the first digit of that block. (This type of declarative
required explicit specification of when to switch between the representation has been used and independently validated in pre-
driving and dialing tasks. For present purposes, we began with the vious models of list memory, Anderson et al., 1998, and analogy,
driver model used in the previous simulations (driving and Salvucci & Anderson, 2001.) Meanwhile, the driving thread per-
choice)—that is, the driver model served as one thread of the full forms its processing, ending with a check of vehicle stability. If the
model, just as in the previous simulations. All parameters were vehicle is not sufficiently stable, the driving thread maintains
kept constant for the current simulations, except that we reesti- control and iterates until stable; if the vehicle is stable, the dialing
mated the stability scaling factor (2.5) again to adjust for different thread intercedes. The dialing thread then locates the desired digit
simulation conditions. Next, we took the original dialing models to enter and moves the hand to this location while visually encod-
Driving Dial
Procedural: Find road near point (waiting until procedural is free)
(waiting until procedural is free) Procedural: Retrieve block
Procedural: Find road far point
Procedural: Steer & accelerate Declarative: Retrieve block
Procedural: Note block

Procedural: Retrieve first digit
Procedural: Note vehicle is stable

Declarative: Retrieve digit
Procedural: Find digit button

(wait for control delay)
Procedural: Press digit button
Proc: Retrieve
Manual: Visual: next digit
Press button Guide press
Declarative:
Retrieve next digit
Proc: Find button,

(waiting until visual is free) or do next block...
Procedural: Press digit button

Proc: Retrieve
Manual: Visual: next digit
Press button Guide press
= Manual: Steer; Pedal:Acc/Decelerate;

Figure 15. Driving– dialing study: Model timeline.
ing the object. In parallel with the button press, the thread retrieves task of driving. The model captures this effect in that although
the next digit to enter. When there are no more digits in the block, dialing time slows down slightly because of the driving task, the
the dialing thread reiterates block retrieval, at which point the threaded model is able to interweave the two tasks well enough
driving thread can reacquire the visual resource and control the that the difference is not large. The lateral velocity results in
vehicle until stable. Thus, the driving thread controls the vehicle Figure 16b indicate that, as mentioned, the humans showed sig-
between blocks but not within blocks; a detailed study of the nificant distraction effects for the manual conditions (comparing
response times for individual digits within the phone number results for driving only to these conditions) and no significant
(Salvucci, 2005) suggests that people indeed exhibit this same effect for the voice conditions. The model also captures this effect,
pattern of behavior. primarily because the manual dialing tasks require visual attention
Figure 16 shows two results from the model simulations and and thus must share this resource with the driving task, creating
human data: (a) the total time needed to complete the dialing task, interference and reduced performance.
both alone and while driving (R2 ⫽ .99, RMSE ⫽ .49), and (b) the Driving and the sentence-span task. Although most studies of
vehicle’s lateral (side-to-side) velocity, an indicator of vehicle driving and phone dialing focus on the contention of resources
stability and a common measure of driver performance (R2 ⫽ .96, (primarily visual) between the tasks, it has also been shown that
RMSE ⫽ .02). The dialing times in Figure 16a show that the even primarily cognitive tasks—for example, conversing with
speed–manual condition is the fastest and full–voice the slowest, another person over the phone— can have detrimental effects on
both for dialing only and while driving. Perhaps surprisingly, the driving performance (see, e.g., Strayer & Johnston, 2001). We now
human drivers required very little extra time to dial with the added examine this issue of cognitive distraction by modeling data re-
Human Model
10
Dialing Only Dialing Only
9
While Driving While Driving
8
7
Dialing Time (s)
6
5
4
3
2
1
0
Full-Manual Speed-Manual Full-Voice Speed-Voice Full-Manual Speed-Manual Full-Voice Speed-Voice
Human Model
.3
Lateral Velocity (m/s)
.2
.1
.0
Driving Full- Speed- Full- Speed- Driving Full- Speed- Full- Speed-
Only Manual Manual Voice Voice Only Manual Manual Voice Voice
Figure 16. Driving-dialing study: Data and model results for dialing time (a) and lateral velocity (b). Error bars
represent the standard errors.
ported by Alm and Nilsson (1995) on drivers performing an involves two difficult activities, namely judging of sentence sen-
intensive cognitive sentence-span task; this model represents a sibility and memorization (and rehearsal) of final words. When
generalization and threaded version of an earlier model for these combined with driving, the task puts a substantial cognitive load
data (Salvucci, 2002). By modeling this task, we aim to demon- on drivers as they attempt to integrate the tasks.
strate that threaded cognition accounts not only for effects of Our model for this task reused components from two existing
resource contention for perceptual and motor resources (as in the models. First, because the driving task is in essence the same as
previous model) but also for effects of contention for cognitive that in the driving and choice domain (i.e., car following and
processing—that is, the contention that arises when executing braking for a lead vehicle), we imported that model directly for use
multiple threads. in the current model, with no changes to rules or parameters. To
The driving task in the Alm and Nilsson (1995) study was a model the sentence-span task, we based our work on that of Lovett,
car-following task in which the lead vehicle would sometimes Daily, and Reder (2000), who modeled a closely related task called
brake suddenly, thus requiring the driver to react and brake in the MODS (modified digit span) task in which people read strings
response. The secondary cognitive task, the sentence-span task, of letters and digits while memorizing final digits for later recall.
was intended as a surrogate for the cognitive load of an intense Because the task and architectural version differed from this ex-
conversation (given that real-life conversations would be difficult isting model, we could not import the model directly; however, the
to control for cognitive load across individuals). The sentence-span original model provided three critical components that were reused
task (see also Daneman & Carpenter, 1980) involved two stages here: (a) the positional representation used to encode memorized
that included the processing of sentences and the recall of words in items, (b) production rules that perform rehearsal of memorized
these sentences. In the first stage, drivers listened to five sentences items, and (c) the basic structure of the production rules that
of the form X does Y—for instance, The boy brushed his teeth or retrieve and report the items in sequence. These rules were then
The train bought a newspaper. After each of these sentences, extended to press a button to start the task, to encode a sentence
drivers reported whether the statement was generally sensible. In and decide whether it was sensible, and to hear and speak words
the second stage, drivers were asked to state the last word of each rather than read and type characters (as in the original MODS
sentence in the order in which they were presented. For instance, model). The process of confirming whether the sentence is sensible
for the sentences The boy brushed his teeth and The train bought was not modeled in any detail, but rather the model simply
a newspaper, the driver would report “yes” and “no” after each assumes that this process happens during the listening productions
sentence (respectively) and would then report the memorized list and signals a confirmation by firing a confirmation rule. In addi-
“teeth,” “newspaper,” and so on. The sentence-span task itself tion, the model assumes that each sentence component (subject,
verb, object) requires 1 s of speech time with a one-syllable pause sentence-span task. First, they found a significant effect of the
(150 ms) in between components. Results were compiled from cognitive task on braking performance as measured by the reaction
three simulation runs with four trials each while driving for ap- time to the lead-vehicle braking event; specifically, they found that
proximately 5 min. Overall, the sentence-span model creates a in the presence of the cognitive task, brake reaction time increased
heavy cognitive load: During much of the slack time in between by 0.56 s. The threaded cognition model of this combined task
encoding auditory stimuli or speaking responses, it rehearses the produced a very similar increase of 0.54 s. The model did not
item list through repeated retrievals of the list, going one by one capture the exact quantitative numbers in the experiment, however:
through all elements of the list and then iterating. Alm and Nilsson reported brake reaction times of approximately
Figure 17 illustrates a segment of a sample process timeline 1.6 s and 2.2 s for the driving-only and with-secondary-task
taken while the model executes both the driving and cognitive conditions (respectively), whereas the model produced times of
tasks together. Throughout the timeline, the model interleaves the 1.00 s and 1.54 s (respectively). Note that our model’s prediction
driving thread with the sentence-span thread. In particular, at the of 1.00 s in the driving-only condition comes directly from the
start of the timeline, the sentence-span thread performs a rehearsal driving and choice model, which fit the Levy et al. (2006) data
retrieval of a list item and then proceeds to check for the presence very well; however, Alm and Nilsson’s time of 1.6 s is quite a bit
of a new audio stimulus (i.e., the spoken sentences streaming in). higher for essentially the same task, and we have no explanation
When a new stimulus is heard (indicated by the star in Figure 17), for this variance. Nevertheless, the model captures the most im-
the sentence-span thread encodes the stimulus and notes it in its portant aspect of their data, namely the magnitude of the perfor-
declarative structure of the current sentence. In this case, the mance decrement when driving while performing the cognitive
stimulus represented the end of the sentence, and thus the thread task.
initiates the required spoken response to the sentence (i.e., the Along with this result, Alm and Nilsson (1995) predicted that
sense judgment). As the vocal resource produces this response, the the cognitive task would also significantly affect drivers’ ability to
retrieval finally completes—the long delay indicating a new de- maintain a central position on the roadway—that is, they predicted
clarative chunk with relatively few rehearsals—and the thread an effect on vehicle position similar to that observed in many
notes the completion in preparation for the next rehearsal retrieval. studies of driving and phone dialing. However, the empirical data
All the while, the driving thread continues its processing, but the did not show such an effect; Alm and Nilsson did not report
many rule firings required by the sentence-span thread result in specific numbers, only that their statistical analysis yielded no
longer intervals between driving updates (roughly twice as long as significant effect of the presence of the cognitive task on lateral
the normal interval). As is evident in Figure 17, the primary position. We analyzed the model behavior for effects of lateral
bottleneck in this processing is the procedural resource, because position by computing the lateral deviation for both conditions as
each thread requires frequent procedural rule firings and thus the RMSE between the vehicle’s position and the lane center (a
contention for this resource remains high throughout the process. common measure of driver performance). For this measure, the
The Alm and Nilsson (1995) study reported two significant model produced a lateral deviation of 0.19 m for the driving-only
findings from their analysis of cognitive distraction from the condition and 0.18 m for the with-secondary-task condition, thus
finding essentially no effect of secondary task and replicating the
findings of Alm and Nilsson’s study. The lack of effect on lateral
Driving Sentence-Span Task deviation is especially interesting given the presence of the effect
Procedural: Find road near point (waiting until procedural is free) for brake reaction time. We attribute this result to the fact that
(waiting until procedural is free)
steering updates become, in essence, elongated and slightly less
Procedural: Rehearse by retrieval
frequent during cognitive distraction but still occur at regular
Procedural: Find road far point (waiting)
intervals; thus, the reduced frequency and longer updates translate
(waiting until procedural is free) Procedural: Listen to additional time for reacting to a braking event, but the regularity
Procedural: Steer & accelerate (waiting) of the steering updates (as filtered through vehicle dynamics) does
(waiting until procedural is free) Procedural: Listen not significantly affect the ability to maintain a central lane posi-
Procedural: Note vehicle is stable Declarative: (waiting) tion. (This regularity in the cognitive distraction task is much like
Retrieve that in the voice dialing tasks in the driving and dialing domain,
(waiting until procedural is free) memorized Procedural: Encode word
list item both producing no significant lateral effects, in contrast to the long
Procedural: Wait for control delay
Aural: Encode word interruptions in the manual dialing tasks that do indeed result in
Procedural: Find road near point significant lateral effects.)
(waiting until procedural is free) Procedural: Note word We should reemphasize the importance of these simulations
Procedural: Find road far point (waiting) with respect to the prediction of contention for cognitive process-
(waiting until procedural is free) Procedural: Respond
ing resources. Unlike all the modeled tasks we have seen thus far,
(waiting)
the combined driving and sentence-span task includes essentially
Procedural: Steer & accelerate
Vocal:
Respond to
no contention for perceptual and motor resources but includes a
(waiting until procedural is free) Procedural: Note retrieval sentence secondary task with a high cognitive load. Thus, the cognitive
Procedural: Note vehicle is stable (waiting) distraction that occurs as a result of the secondary task primarily
arises because of contention for cognitive processing. Some cog-
= Manual: Steer; Pedal: Acc/Decelerate; = Next word can be heard
nitive architectures, most saliently EPIC (Meyer & Kieras, 1997a),
posit a fully parallel cognitive processor with no constraints on
Figure 17. Driving-sentence-span study: Model timeline. processing resources. Such an architecture would thus predict no
effects of multitasking performance in this task, because both tasks origins of the models and the changes incorporated for our
could execute comfortably in parallel with no perceptual or motor simulations.
contention to constrain them. In contrast, threaded cognition and
the ACT-R cognitive architecture posit a resource-bounded cog- General Discussion
nitive processor that nicely captures the observed performance
decrements that arise in cognitive distraction tasks. Threaded cognition provides a theoretical and computational
framework for reasoning about and predicting multitasking behav-
Summary of Model Simulations ior. The model simulations emphasize the theory’s computational
aspects, namely its ability to generate detailed predictions of
The model simulations illustrate the many types of interference behavior across a variety of tasks. In this discussion, we expound
that can arise in multitasking behavior and demonstrate that on the broader theoretical implications of threaded cognition for
threaded cognition captures numerous aspects of this behavior. concurrent multitasking and related types of multitasking behavior.
The various domains each emphasize different sources of process-
ing interference. Although all the task domains include some Theoretical Implications for Concurrent Multitasking
degree of cognitive interference resulting from the reliance on
cognition for procedural rule firings, the procedural resource The first implication of our theory of threaded cognition is that
serves only as a primary source of interference in the driving/ concurrent multitasking does not require supervisory or executive
sentence-span domain, whereas it serves as a secondary source of processes to manage and schedule multiple task processes. Any
interference in the dual-choice and driving-choice domains. The approach that requires such processes must specify either how
dual-choice domain emphasizes interference in declarative mem- executive processes could be acquired for particular sets of tasks or
ory retrieval, primarily in the early stages of learning before such how a general executive or supervisory mechanism would operate
retrievals drop out. The tracking-choice and driving-dialing do- over all sets of tasks; no previous theory has successfully provided
mains emphasize interference effects in perceptual processing, and such an explanation. In contrast, threaded cognition allows multi-
the tracking-choice domain also demonstrates interference effects tasking to emerge from the interaction of autonomous threads and
in motor processing. In addition, all the domain models demon- a threading mechanism for basic resource acquisition and conflict
strate varying amounts of model reuse and transfer both from resolution. Our modeling results demonstrate that this straightfor-
earlier work and within the current work; Table 2 summarizes the ward mechanism suffices to account for behavior across a range of
Table 2
Summary of Model Origins and Parameter Settings
Domain Origins and parameter settings
Dual choice Choice model based on Taatgen (2005)

No parameters estimated; all values taken from Taatgen (2005) and Byrne & Anderson (2001)
Tracking and choice Tracking model newly developed
No parameters estimated; all parameters left at default values
Choice model adapted from dual-choice study
Model changed to identify arrow stimuli instead of previous stimuli
No parameters estimated
Reading and dictation Reading model based on Lewis & Vasishth (2005)
Model implementation a simplified version of original (which would require too much time to run for
simulated days)
Set 1 parameter, latency factor, so that the timing of all model actions is the same as in the original model
Dictation model newly developed
Estimated 2 parameters—motor action and procedural learning speed
Driving and choice Driving model based on Salvucci (2005)
Model implementation changed to new version of ACT-R
Model changed to break out of normal delay due to braking task
Estimated 3 parameters—steering scaling factor, stability scaling factor, and accelerator-to-brake time
Choice model adapted from dual choice study
Model changed to accept dual-instance stimulus
Estimated 1 parameter—voice recognition delay
Driving and dialing Driving model taken from driving and choice study
Model changed to remove braking task (not applicable here)
Estimated 1 parameter—stability scaling factor
Dialing model taken from Salvucci (2001b)
Assumed that digit is 1 syllable and phrases are 3 syllables
No new parameters estimated
Driving and sentence-span task Driving model taken from driving and choice study:
No new parameters estimated
Sentence-span model derived from Lovett et al. (2000)
Model uses representation and some rules from earlier model
multitasking domains, including accounts of when behavior does planning or strategizing about multitasking performance; it simply
and does not exhibit multitasking interference. states that any such higher level processing must be done by a
The second implication of the theory is that both procedural and cognitive thread rather than a specialized parallel executive pro-
declarative processes can be sources of dual-task interference and cess and thus must share resources and perhaps experience inter-
that declarative processes are an especially prevalent source of ference from other concurrent threads. For example, De Jong
interference in the early stages of learning. The procedural re- (1995) found that in a PRP-like dual-choice task, people respond
source, as the nexus of processing in our theory, is always present faster when they can predict which choice task will come first, and
as a potential source of interference because its processing is they respond slower when their expectation is incorrect. Although
required to collect results from and initiate processes in other threaded cognition itself would not account for this result, such
resources. The declarative resource, on the other hand, is most behavior can arise from additional rules that plan for one task over
active in early stages of learning because of dependence on mem- the other (similar to EPIC executive processes forced to run on a
orized task instructions. Over time, however, this dependence serial procedural processor). In essence, the task would be learned
reduces and drops out because of the gradual translation of declar- as a single thread that performs both tasks in a desired order. In
ative to procedural skill. Although the presence of declarative fact, the participants in De Jong’s study experienced the dual-task
interference has been previously explored (e.g., Carrier & Pashler, condition immediately with no single-task practice and were ex-
1995), threaded cognition makes explicit the roles of the two plicitly instructed to pay attention to task ordering. As another
resources and how these roles change over time with practice. example, Logan and Burkell (1986) introduced stop and change
The third implication is that the representation of component variants of the dual-task paradigm in which a stimulus in one task
task skills is simple and parsimonious, free from task-specific affects the response to the other task; again, these variants pre-
knowledge that dictates when and how task switching should sented two closely tied tasks and thus also biased participants to
occur. With such representations, task skills can be easily inte- more of a single-thread representation. In both examples, the
grated or separated, allowing for flexible combinations of single- resulting representations include rules for higher level planning
and dual-task behavior. This aspect of the theory meshes well with and reasoning, but no truly concurrent processing akin to what
the typical methodology for multitasking studies in which compo- threaded cognition is intended to address.
nent task skills are learned independently: Most empirical studies Another aspect of such behavior relates to a person’s ability to
ranging from simple dual-task experiments (e.g., Byrne & Ander- modulate their performance on one task with respect to another in
son, 2001; Schumacher et al., 1999) to complex studies of sec- a multitasking context. We often conceptualize multiple tasks as
ondary tasks while driving (e.g., Alm & Nilsson, 1995; Salvucci, having priorities in which one task should receive a larger propor-
2001b) provide participants with a warm-up period in which one or tion of processing time than another. However, defining such
both tasks can be practiced independently of other tasks. Thus, we priorities explicitly can be a difficult endeavor; for instance, in our
expect that skill knowledge acquired during this practice time driving and phone dialing example, driving should clearly have the
remains uncluttered, without the additional baggage of task- much higher priority and yet people still dial phones while driv-
switching knowledge. This being the case, threaded cognition ing— how can we specify and account for such behavior in our
provides the mechanism with which to execute such skills together models? Rather than defining priorities per se, our approach allows
in a multitasking context. task representations to incorporate, to some extent, their own
In fact, a fourth implication of our theory is that, because of the requirements for processing time. For example, when our driver
absence of task-specific executive knowledge, practicing two tasks model notes that the vehicle is stable, it allows for a short delay
concurrently results in the same performance as practicing the two before more vehicle control must be performed; by adjusting this
tasks independently. This implication may seem surprising but has delay as external demands change, the driver model can essentially
recently been supported by Ruthruff, Van Selst, Johnston, and adapt its processing demands over time (see Salvucci et al., 2006).
Remington (2006) in a study of practice effects in the dual-choice Such adaptations allow threaded cognition to account for differing
task domain. Thus, threaded cognition accounts for the entire span processing trade-offs between two concurrent tasks, such as trade-
of learning, from the initial stages of instruction learning to the offs characterized as points along a performance operating char-
latest asymptotic stages of highly optimized behavior—an impor- acteristic curve (see, e.g., Navon & Gopher, 1979; Norman &
tant step toward replacing the oft-derided homunculus of control Bobrow, 1975).
(Altmann, 2003; Logan, 2003; Monsell & Driver, 2000) with a Threaded cognition is also not meant to account for multitasking
fully specified computational framework. at the minute-to-hour-long time scale, such as multitasking be-
tween two long-term work projects but focusing on one for an
Implications for Other Aspects of Multitasking extended period before switching to the other. For our purposes,
this type of multitasking is better characterized as performing
Although our theory applies across a range of interesting exper- sequential tasks with occasional interruptions. These areas of study
imental and real-world domains, threaded cognition focuses on do not go against the threaded approach presented here; indeed, we
nondeliberative multitasking in which concurrent tasks are per- believe that these theories and models nicely complement the types
formed at the subsecond to second time scale. In contrast, the of multitasking addressed by threaded cognition and that they can
theory is not meant to capture explicit deliberative multitasking as very likely fit well in the context of a threaded approach.
might arise in some contexts. For example, as discussed earlier, we To illustrate our point, we can consider two bodies of literature
do not aim to capture the planning-centered, nongreedy types of in particular through the lens of a threaded approach. First, the
multitasking that other efforts have addressed (e.g., Freed, 1998; sizable literature on task switching (see, e.g., Altmann & Gray,
Howes et al., 2004). Nevertheless, the theory does not preclude 2002; Rogers & Monsell, 1995; Sohn & Anderson, 2001) has
explored various aspects of switch costs when performing succes- will generalize to an increasingly diverse array of multitasking
sive choice tasks—for example, alternating between two distinct domains.
choice tasks. For such a domain, the two alternating tasks do not
fall under the context of a threaded approach, because threading is
intended to capture simultaneous execution of tasks rather than 2
See http://act-r.psy.cmu.edu/
separate, sequential execution. In this case, a model would be
required to explicitly change the control state from the current task
References
to the other task, likely requiring a memory retrieval or other costly
cognitive function. There have been computational accounts for Alm, H., & Nilsson, L. (1994). Changes in driver behaviour as a function
these switch costs based on theories of memory and recall (e.g., of hands-free mobile phones—A simulator study. Accident Analysis &
Altmann, 2002; Sohn & Anderson, 2001). Such an account, ex- Prevention, 26, 441– 451.
pressed in the ACT-R framework, can also be expressed in a Alm, H., & Nilsson, L. (1995). The effects of a mobile telephone task on
threaded framework by simply assuming that the model acts as a driver behaviour in a car following situation. Accident Analysis &
Prevention, 27, 707–715.
single thread, explicitly redirecting itself to accomplish alternating
Altmann, E. M. (2002). Functional decay of memory for tasks. Psycho-
tasks represented by different control states. logical Research, 66, 287–297.
A second interesting body of literature addresses the issue of Altmann, E. M. (2003). Task switching and the pied homunculus: Where
interruption. For instance, Trafton, Altmann, Brock, and Mintz are we being led? Trends in Cognitive Sciences, 7, 340 –341.
(2003) explored people’s behavior during the interruption lag, or Altmann, E. M., & Gray, W. D. (2002). Forgetting to remember: The
the time interval between an alert of an interruption (e.g., a phone functional relationship of decay and interference. Psychological Science,
ringing) and the actual interrupting task (e.g., the phone conver- 13, 27–33.
sation). They found that, when interrupted, people used the inter- Altmann, E. M., & Trafton, J. G. (2002). Memory for goals: An activation-
ruption lag to prepare for later resumption of the interrupted task based model. Cognitive Science, 26, 39 – 83.
(see also Altmann & Trafton, 2002). In particular, they posit that Anderson, J. R., Bothell, D., Byrne, M. D., Douglass, S., Lebiere, C., &
Qin, Y. (2004). An integrated theory of the mind. Psychological Review,
people use prospective goal encoding as well as retrospective goal
111, 1036 –1060.
rehearsal to maintain a high activation on the interrupted goal, Anderson, J. R., Bothell, D., Lebiere, C., & Matessa, M. (1998). An
facilitating later retrieval. Such a model could be neatly expressed integrated theory of list memory. Journal of Memory and Language, 38,
in a threaded approach: When the alert occurs, a thread is created 341–380.
that maintains high memory activation by repeated retrievals of the Anderson, J. R., Taatgen, N. A., & Byrne, M. D. (2005). Learning to
control state (much like our model of the sentence-span task in its achieve perfect time sharing: Architectural implications of Hazeltine,
rehearsal of the word list). During the interruption lag and even Teague, & Ivry (2002). Journal of Experimental Psychology: Human
during the interrupting task, such a thread can occasionally per- Perception and Performance, 31, 749 –761.
form retrieval and ensure that, upon completion of the interrupting Baddeley, A. D. (1986). Working memory. Oxford, England: Oxford Uni-
task, the original task can be restarted by retrieval of its control versity Press.
Baddeley, A., Lewis, V., Eldridge, M., & Thomson, N. (1984). Attention
state. Another interpretation might be that the thread of the original
and retrieval from long-term memory. Journal of Experimental Psychol-
task remains during the interrupting task, but the context of the ogy: General, 113, 518 –540.
original goal is forgotten, requiring retrieval of this context before Broadbent, D. E. (1958). Perception and communication. Elmsford, NY:
resumption. Whatever the approach, the theoretical constructs of Pergamon Press.
Trafton et al.’s theory on interruption and resumption remain Brookhuis, K. A., De Vries, G., & De Waard, D. (1991). The effects of
wholly intact, but threaded cognition provides a straightforward mobile telephoning on driving performance. Accident Analysis & Pre-
method to accomplish their posited goal rehearsal—in essence, as vention, 23, 309 –316.
a second thread along with the interrupting task thread. Byrne, M. D., & Anderson, J. R. (2001). Serial modules in parallel: The
Looking ahead, the ultimate test of threaded cognition arises in psychological refractory period and perfect time-sharing. Psychological
the application of the theory across a wide range of task domains. Review, 108, 847– 869.
Carrier, L. M., & Pashler, H. (1995). Attentional limits in memory re-
The model simulations presented here represent a first step at
trieval. Journal of Experimental Psychology: Learning, Memory, and
demonstrating the theory’s breadth and generality using an illus- Cognition, 21, 1339 –1348.
trative representative sampling of domains. As a next step, one Chong, R. S. (1998). Modeling dual-task performance improvement: Cast-
might consider the vast space of domains for which cognitive ing executive process knowledge acquisition as strategy refinement
architectural models have been developed and explore the theory’s (Tech. Rep. No. CSE-TR-378 –98). Ann Arbor: University of Michigan,
predictions for when two or more task domains are combined into Department of Computer Science and Engineering.
a single concurrent task. For example, the ACT-R Web site2 Cooper, R., & Shallice, T. (2000). Contention scheduling and the control of
contains over 600 publications describing computational models routine activities. Cognitive Neuropsychology, 17, 297–338.
covering over 50 task domains; one might imagine combining any Craik, K. J. W. (1948). Theory of the human operator in control systems:
two tasks into a single concurrent task (e.g., lexical processing and II. Man as an element in a control system. British Journal of Psychology,
38, 142–148.
spatial navigation, cognitive arithmetic and game playing, etc.),
Daneman, M., & Carpenter, P. A. (1980). Individual differences in working
generating predictions of the concurrent task by simply integrating memory and reading. Journal of Verbal Learning and Verbal Behavior,
the task models as threads and then comparing these predictions to 19, 450 – 466.
(perhaps newly collected) empirical data on the concurrent task. De Jong, R. (1995). The role of preparation in overlapping-task perfor-
Thus, threaded cognition is an eminently testable theory of mul- mance. Quarterly Journal of Experimental Psychology: Human Exper-
titasking behavior, and we look forward to exploring how well it imental Psychology, 48(A), 2–25.
Freed, M. (1998). Managing multiple tasks in complex, dynamic environ- Liu, Y. (1996). Queuing network modeling of elementary mental pro-
ments. In Proceedings of the 1998 National Conference on Artificial cesses. Psychological Review, 103, 116 –136.
Intelligence (pp. 921–927). Menlo Park, CA: AAAI Press. Liu, Y., Feyen, R., & Tsimhoni, O. (2005). Queuing Network-Model
Gluck, K. A., & Pew, R. W. (Eds.). (2005). Modeling human behavior with Human Processor (QN-MHP): A computational architecture for multi-
integrated cognitive architectures: Comparison, evaluation, and valida- task performance in human-machine systems. ACM Transactions on
tion. Mahwah, NJ: Erlbaum. Computer-Human Interaction, 13, 37–70.
Gray, W. D. (Ed.) (2007). Integrated models of cognitive systems. Oxford, Logan, G. D. (2003). Executive control of thought and action: In search of
England: Oxford University Press. the wild homunculus. Current Directions in Psychological Science, 12,
Graybiel, A. M., & Kimura, M. (1995). Adaptive neural networks in the 45– 48.
basal ganglia. In J. C. Houk, J. L. Davis, & D. G. Beiser (Eds.), Models Logan, G. D., & Burkell, J. (1986). Dependence and independence in
of information processing in the basal ganglia (pp. 103–116). Cam- responding to double stimulation: Comparison of stop, change, and
bridge, MA: MIT Press. dual-task paradigms. Journal of Experimental Psychology: Human Per-
Hazeltine, E., Teague, D., & Ivry, R. B. (2002). Simultaneous dual-task ception and Performance, 12, 549 –563.
performance reveals parallel response selection after practice. Journal of Lovett, M. C., Daily, L. Z., & Reder, L. M. (2000). A source activation
Experimental Psychology: Human Perception and Performance, 28, theory of working memory: Cross-task prediction of performance in
527–545. ACT-R. Journal of Cognitive Systems Research, 1, 99 –118.
Howes, A., Vera, A., & Lewis, R. L. (2007). Bounding rational analysis: Martin-Emerson, R., & Wickens, C. D. (1992). The vertical visual field and
Constraints on asymptotic performance. In W. D. Gray (Ed.), Integrated implications for the head-up display. In Proceedings of the Thirty-Sixth
models of cognitive systems (pp. 403– 413). Oxford, England: Oxford Annual Symposium of the Human Factors Society (pp. 1408 –1412).
University Press. Santa Monica, CA: Human Factors Society.
Howes, A., Vera, A., Lewis, R. L., & McCurdy, M. (2004). Cognitive McKnight, A. J., & McKnight, A. S. (1993). The effect of cellular phone
constraint modeling: A formal approach to supporting reasoning about use upon driver attention. Accident Analysis & Prevention, 25, 259 –265.
behavior. In K. Forbus, D. Gentner, & T. Reiger (Eds.), Proceedings of Meyer, D. E., & Kieras, D. E. (1997a). A computational theory of exec-
the 26th Annual Conference of the Cognitive Science Society (pp. 595– utive cognitive processes and multiple-task performance: Part 1. Basic
600). Hillsdale, NJ: Erlbaum. mechanisms. Psychological Review, 104, 3– 65.
John, B. E. (1996). TYPIST: A theory of performance in skilled typing. Meyer, D. E., & Kieras, D. E. (1997b). A computational theory of exec-
Human-Computer Interaction, 11, 321–355. utive cognitive processes and multiple-task performance: Part 2. Ac-
Johnston, J. C., McCann, R. S., & Remington, R. W. (1995). Chronometric counts of psychological refractory period phenomena. Psychological
evidence for two types of attention. Psychological Science, 6, 365–369. Review, 104, 749 –791.
Jones, R. M., Laird, J. E., Nielsen, P. E., Coulter, K., Kenny, P., & Koss, Monsell, S., & Driver, J. (2000). Banishing the control homunculus. In S.
F. (1999). Automated intelligent pilots for combat flight simulation. AI Monsell & J. Driver (Eds.), Control of cognitive processes: Attention
Magazine, 20, 27– 42. and performance XVIII (pp. 3–32). Cambridge, MA: MIT Press.
Just, M. A., Carpenter, P. A., & Varma, S. (1999). Computational modeling Navon, D., & Gopher, D. (1979). On the economy of the human-processing
of high-level cognition and brain function. Human Brain Mapping, 8, system. Psychological Review, 86, 214 –255.
128 –136. Newell, A., & Simon, H. (1972). Human problem solving. Englewood
Keele, S. W. (1973). Attention and human performance. Pacific Palisades, Cliffs, NJ: Prentice-Hall.
CA: Goodyear. Norman, D. A., & Bobrow, D. G. (1975). On data-limited and resource-
Kieras, D. E. (2007). Control of cognition. In W. D. Gray (Ed.), Integrated limited processes. Cognitive Psychology, 7, 44 – 64.
models of cognitive systems (pp. 327–355). Oxford, England: Oxford Norman, D. A., & Shallice, T. (1986). Attention to action: Willed and
University Press. automatic control of behavior. In R. J. Davidson, G. E. Schwartz, & D.
Kieras, D. E., Meyer, D. E., Ballas, J. A., & Lauber, E. J. (2000). Modern Shapiro (Eds.), Consciousness and self-regulation (pp. 1–18). New
computational perspectives on executive mental processes and cognitive York: Plenum.
control: Where to from here? In S. Monsell & J. Driver (Eds.), Control O’Reilly, R. C., & Frank, M. J. (2006). Making working memory work: A
of cognitive processes: Attention and performance XVIII (pp. 681–712). computational model of learning in the prefrontal cortex and basal
Cambridge, MA: MIT Press. ganglia. Neural Computation, 18, 283–328.
Kucera, H., & Francis, W. N. (1967). Computational analysis of present- Pashler, H. (1994). Dual-task interference in simple tasks: Data and theory.
day American English. Providence, RI: Brown University Press. Psychological Bulletin, 116, 220 –244.
Laird, J. E., Newell, A., & Rosenbloom, P. S. (1987). Soar: An architecture Pashler, H., & Johnston, J. C. (1998). Attentional limitations in dual-task
for general intelligence. Artificial Intelligence, 33, 1– 64. performance. In H. Pashler (Ed.), Attention (pp. 155–189). Hove, En-
Lallement, Y., & John, B. E. (1998). Cognitive architecture and modeling gland: Taylor & Francis.
idiom: An examination of three models of the Wickens task. In M. A. Reichle, E. D., Pollatsek, A., Fisher, D. L., & Rayner, K. (1998). Toward
Gernsbacher & S. J. Derry (Eds.), Proceedings of the Twentieth Annual a model of eye movement control in reading. Psychological Review, 105,
Conference of the Cognitive Science Society (pp. 597– 602). Hillsdale, 125–157.
NJ: Erlbaum. Rogers, R. D., & Monsell, S. (1995). Costs of a predictable switch between
Lee, J. D., McGehee, D. V., Brown, T. L., & Reyes, M. L. (2002). simple cognitive tasks. Journal of Experimental Psychology: General,
Collision warning timing, driver distraction, and driver response to 124, 207–231.
imminent rear-end collisions in a high fidelity driving simulator. Human Rohrer, D., & Pashler, H. E. (2003). Concurrent task effects on memory
Factors, 44, 314 –334. retrieval. Psychonomic Bulletin & Review, 10, 96 –103.
Levy, J., Pashler, H., & Boer, E. (2006). Central interference in driving: Is Rumelhart, D. E. (1980). Schemata: The building blocks of cognition. In
there any stopping the psychological refractory period? Psychological R. J. Spiro, B. Bruce, & W. F. Brewer (Eds.), Theoretical issues in
Science, 17, 228 –235. reading and comprehension (pp. 33–58). Hillsdale, NJ: Erlbaum.
Lewis, R. L., & Vasishth, S. (2005). An activation-based model of sentence Rumelhart, D. E., Smolensky, P., McClelland, J. L., & Hinton, G. E.
processing as skilled memory retrieval. Cognitive Science, 29, 375– 419. (1986). Schemata and sequential thought processes in PDP models. In
D. E. Rumelhart & J. L. McClelland (Eds.), Parallel distributed pro- dual-task performance: Uncorking the central cognitive bottleneck. Psy-
cessing (Vol. 2, pp. 7–57). Cambridge, MA: MIT Press. chological Science, 12, 101–108.
Ruthruff, E., Van Selst, M., Johnston, J. C., & Remington, R. (2006). How Sohn, M.-H., & Anderson, J. R. (2001). Task preparation and task repeti-
does practice reduce dual-task interference: Integration, automatization, tion: Two-component model of task switching. Journal of Experimental
or just stage-shortening? Psychological Research, 70, 125–142. Psychology: General, 130, 764 –778.
Salvucci, D. D. (2001a). An integrated model of eye movements and visual Spelke, E., Hirst, W., & Neisser, U. (1976). Skills of divided attention.
encoding. Cognitive Systems Research, 1, 201–220. Cognition, 4, 215–230.
Salvucci, D. D. (2001b). Predicting the effects of in-car interface use on Strayer, D. L., & Johnston, W. A. (2001). Driven to distraction: Dual-task
driver performance: An integrated model approach. International Jour- studies of simulated driving and conversing on a cellular telephone.
nal of Human-Computer Studies, 55, 85–107. Psychological Science, 12, 462– 466.
Salvucci, D. D. (2002). Modeling driver distraction from cognitive tasks. In Taatgen, N. A. (2005). Modeling parallelization and flexibility improve-
Proceedings of the 24th Annual Conference of the Cognitive Science ments in skill acquisition: From dual tasks to complex dynamic skills.
Society (pp. 792–797). Mahwah, NJ: Erlbaum. Cognitive Science, 29, 421– 455.
Salvucci, D. D. (2005). A multitasking general executive for compound Taatgen, N. A., & Lee, F. J. (2003). Production compilation: A simple
mechanism to model complex skill acquisition. Human Factors, 45,
continuous tasks. Cognitive Science, 29, 457– 492.
61–76.
Salvucci, D. D. (2006). Modeling driver behavior in a cognitive architec-
Taatgen, N. A., van Rijn, H., & Anderson, J. R. (2007). An integrated
ture. Human Factors, 48, 362–380.
theory of prospective time interval estimation: The role of cognition,
Salvucci, D. D., & Anderson, J. R. (2001). Integrating analogical mapping
attention and learning. Psychological Review, 114, 577–598.
and general problem solving: The path-mapping theory. Cognitive Sci-
Telford, C. W. (1931). The refractory phase of voluntary and associative
ence, 25, 67–110.
response. Journal of Experimental Psychology, 14, 1–35.
Salvucci, D. D., & Gray, R. (2004). A two-point visual control model of Trafton, J. G., Altmann, E. M., Brock, D. P., & Mintz, F. E. (2003).
steering. Perception, 33, 1233–1248. Preparing to resume an interrupted task: Effects of prospective goal
Salvucci, D. D., & Macuga, K. L. (2002). Predicting the effects of cellular- encoding and retrospective rehearsal. International Journal of Human-
phone dialing on driver performance. Cognitive Systems Research, 3, Computer Studies, 58, 583– 603.
95–102. Welford, A. T. (1952). The “psychological refractory period” and the
Salvucci, D. D., Taatgen, N. A., & Kushleyeva, Y. (2006). Learning when timing of high speed performance: A review and a theory. British
to switch tasks in a dynamic multitasking environment. In Proceedings Journal of Psychology, 43, 2–19.
of the Seventh International Conference on Cognitive Modeling (pp. Wickens, C. D. (1984). Processing resources in attention. In R. Parasura-
268 –273). Trieste, Italy: Edizioni Goliardiche. man, J. Beatty, & R. Davies (Eds.), Varieties of attention (pp. 63–101).
Schumacher, E. H., Lauber, E. J., Glass, J. M., Zurbriggen, E. L., Gmeindl, New York: Wiley.
L., Kieras, D. E., & Meyer, D. E. (1999). Concurrent response-selection Wickens, C. D. (2002). Multiple resources and performance prediction.
processes in dual-task performance: Evidence for adaptive executive Theoretical Issues in Ergonomics Science, 3, 159 –177.
control of task scheduling. Journal of Experimental Psychology: Human
Perception and Performance, 25, 791– 814. Received October 12, 2006
Schumacher, E. H., Seymour, T. L., Glass, J. M., Fencsik, D. E., Lauber, Revision received August 1, 2007
E. J., Kieras, D. E., et al. (2001). Virtually perfect time sharing in Accepted August 2, 2007 䡲

Threaded Cognition Published

Uploaded by

Copyright:

Available Formats

Threaded Cognition Published

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Threaded Cognition Published

Uploaded by

Copyright:

Available Formats

Psychological Review Copyright 2008 by the American Psychological Association

2008, Vol. 115, No. 1, 101–130 0033-295X/08/$12.00 DOI: 10.1037/0033-295X.115.1.101

Threaded Cognition: An Integrated Theory of Concurrent Multitasking

Dario D. Salvucci Niels A. Taatgen

Mixer Batter Pasta Stove

Cook: Preheat oven (waiting until cook is free)

Cook: Place fish in oven (waiting until cook is free)

Cook: Stop mixer

Cook: Strain pasta in colander

Cook: Put batter in oven

Oven: Bake cake

Attend-stimulus central procedural resource is needed at each step to initiate new

Visual-Manual Choice Task

Procedural: Attend stimulus

Visual: Encode stimulus

Procedural: Respond to left stimulus

Manual: Press index finger

Vocal Manual Vocal Manual

Figure 2. Single-choice task. a: Model timeline. b: Storyboard view.

Threaded Processing Assumption: Cognition maintains a set 1

Vis-Man Declarative Vis-Man Declarative

Vocal Manual Vocal Manual

Vis-Man Declarative Declarative

Vocal Manual Vocal Manual

"O--" Low tone

Tracking and Reading and Driving and

Figure 6. Dual-choice study: Experiment 1 data and model results.

0.0 0.1 0.2 0.3 0.4 0.5

Experiment 3: Visual−Manual Experiment 3: Aural−Vocal

Reaction time (s)

Figure 8. Dual-choice study: Experiment 3 data and model results.

Procedural: Attend word

Visual: Encode word

(waiting until procedural is free) Procedural: Attend word

Procedural: Attach syntax Aural: Encode word

Procedural: Retrieve word

Procedural: Do lexical access (waiting until procedural is free)

(waiting) Vision: Procedural: Initiate movement

Procedural: Attach syntax

Procedural: Do lexical access

Procedural: Do lexical access

Procedural: Attach syntax

Procedural: Do lexical access

Procedural: Attach syntax

= Declarative: Lexical access (reading) or retrieve spelling (dictation)

Figure 11. Reading-dictation study: Model timeline.

1.0 vucci, 2005, for parameter descriptions): minor adjustments to the

change to the accelerator-to-brake movement delay (600 ms) to

Reyes, 2002) plus 200-ms motor preparation. The final modifica-

The original model incorporated a delay such that in stable situa-

that in stable situations, continual control is not required, and this

.4 Brake Dual Task Brake Dual Task

Procedural: Find road near point (waiting until procedural is free)

(waiting until procedural is free) Procedural: Retrieve block

Procedural: Find road far point

Procedural: Steer & accelerate Declarative: Retrieve block

Procedural: Note block

Procedural: Note vehicle is stable

Procedural: Find digit button

Proc: Find button,