ACT Theory: Architecture of the Human Mind

Understanding the fundamental mechanisms that govern human intelligence has long stood as one of the central quests of cognitive science. Adaptive Control of Thought theory provides a unified, computationally rigorous framework that elucidates how declarative knowledge transforms into streamlined procedural competence during learning, problem solving, and complex task execution.

Adaptive Control of Thought Theory

1. Concise Definition

The Adaptive Control of Thought theory (commonly abbreviated as ACT, and in its contemporary computational instantiation, ACT-R) is a comprehensive cognitive architecture formulated by John Robert Anderson. It posits that human cognition emerges from the coordinated interaction between two distinct knowledge systems: declarative memory (factual knowledge represented as discrete chunks) and procedural memory (operational skill represented as production rules).

At its core, the theory conceptualizes the human mind as an integrated processing system wherein declarative facts are accessed, manipulated, and converted into automatic behavioral routines through iterative cycles of pattern matching, execution, and empirical reinforcement. The architecture models both the symbolic structures that organize conceptual representations and the subsymbolic mathematical mechanisms that govern memory activation, retrieval latency, and decision utility.

By integrating cognitive modules spanning perceptual processing, motor execution, goal tracking, and long-term storage via limited-capacity buffers, the framework serves simultaneously as a theoretical model of mental functioning and as a validated simulation environment capable of reproducing human behavioral data, learning trajectories, and neural activation patterns with millisecond precision.

2. Etymology & Linguistic Origin

The nomenclature “Adaptive Control of Thought” reflects the cybernetic and information-processing heritage of late-twentieth-century cognitive science. The term “Adaptive” originates from the Latin verb adaptare, signifying the capacity of an organism or system to adjust its internal configurations to meet the demands of an evolving environment. Within Anderson’s framework, adaptiveness refers specifically to the rational optimization of cognitive mechanisms—such that memory retrieval and decision pathways align with the statistical structure of environmental recurrences.

“Control” derives from the Anglo-French contreroller (originally denoting the maintenance of a duplicate register or counter-roll for systematic verification), highlighting the governing mechanisms that steer attention, prioritize concurrent goals, and sequence motor actions. “Thought” stems from the Old English þōht, encompassing mental operations, deliberative reasoning, and ideational synthesis.

The progressive appendances to the acronym chronicle its evolution: “ACT” (Adaptive Control of Thought, 1976), “ACT*” (pronounced “ACT-star”, representing the 1983 integration of generalized production compilation and induction), and “ACT-R” (where the “R” signifies “Rational,” instituted in 1993 to denote the incorporation of rational analysis and Bayesian principles of environmental optimization).

3. Pronunciation & Grammatical Form

The term is formally pronounced as an acronym or initialism depending on the variant:

• ACT: Pronounced phonetically as a single lexical unit: /ækt/ (rhyming with “fact”).
• ACT*: Pronounced as “act-star”: /ækt stɑːr/.
• ACT-R: Pronounced sequentially as “act-are”: /ækt ɑːr/.
• Full Name: Adaptive Control of Thought: /əˈdæp.tɪv kənˈtroʊl əv θɔːt/.

Grammatically, “Adaptive Control of Thought” functions as a singular, non-count proper noun phrase. It routinely modifies other nouns in attributive constructions, such as “ACT-R architecture,” “ACT cognitive model,” or “ACT production system.” When describing the theoretical framework generally, it takes a singular verb agreement (e.g., “Adaptive Control of Thought demonstrates how knowledge proceduralization reduces working memory load”).

4. Detailed Conceptual Explanation

Adaptive Control of Thought theory operates on the premise that all human mental activities—from basic sensory-motor coordination to abstract mathematical deduction—are subserved by invariant cognitive mechanisms operating over task-specific knowledge structures. The central theoretical thesis maintains a strict epistemological separation between “knowing that” (declarative memory) and “knowing how” (procedural memory). Declarative knowledge encompasses facts, episodic encounters, sensory impressions, and relational propositions encoded as symbolic units termed chunks. Each chunk possesses an explicit category type and bounded slots containing specific values (for instance, the arithmetic fact 3 + 4 = 7 or the spatial proposition Paris is the capital of France).

In contrast, procedural knowledge is encapsulated within condition-action rules known as productions. A production rule takes the formal computational syntax of: IF [conditions are met in working memory buffers], THEN [execute specific internal or external actions]. Productions do not possess introspective access to their own rules; instead, they function as dynamic operators that read the current configuration of the mind’s communicative buffers, alter mental contents, request memory retrievals, or dispatch motor directives to external effectors. Thought proceeds through cyclical, discrete 50-millisecond production cycles during which the system evaluates whether the conditional premises of any production rule match the contents of active buffers, selects the optimal rule based on calculated utility, and triggers its consequential action.

A paramount conceptual hallmark of the architecture is its hybrid symbolic-subsymbolic duality. While symbolic entities (chunks and productions) dictate the qualitative topology of cognition, their deployment is continuously modulated by continuous subsymbolic variables. Every declarative chunk possesses a fluctuating level of base-level activation that mirrors its historical frequency and recency of usage, augmented by spreading activation routed from current attentional anchors. The probability of retrieving a chunk, as well as the chronological latency required for that retrieval to occur, is mathematically determined by this cumulative activation value relative to an absolute retrieval threshold.

Concurrently, procedural rules are governed by a subsymbolic utility calculation. When multiple production rules match the identical buffer configuration (a state termed conflict resolution), the architecture does not pick arbitrarily; rather, it computes the expected mathematical utility of each candidate production based on its historical probability of leading to goal achievement relative to the computational and temporal cost expended. Consequently, the architecture captures both the discrete, rule-governed nature of deliberate reasoning and the graded, probabilistic, noise-tolerant characteristics observed in biological learning systems.

Knowledge acquisition in ACT theory is fundamentally evolutionary. When encountering a novel domain, a learner must rely on explicit, declarative instructions stored as chunks. These chunks must be laboriously retrieved into working memory buffers and interpreted through general-purpose, weak-method productions (such as means-ends analysis or analogy). This mode of performance is notably slow, cognitively demanding, and error-prone. As the learner repeatedly solves tasks within the domain, an automatic process called production compilation occurs. This mechanism synthesizes multiple declarative retrievals and interpretive production steps into a singular, highly specialized, domain-specific production rule. Once compiled, this new rule executes directly whenever the environmental conditions arise, bypassing declarative retrieval entirely and freeing executive resources for secondary goals.

5. Historical Development

The evolutionary trajectory of ACT spans more than five decades of theoretical refinement, empirical testing, and computational re-engineering under the direction of John R. Anderson and his associates at Carnegie Mellon University.

The lineage originated in the early 1970s with the development of HAM (Human Associative Memory), co-authored by Anderson and Gordon Bower in 1973. HAM was a pioneering propositional network model designed to simulate how associative representations store linguistic and conceptual information. Although successful in capturing associative memory paradigms, HAM lacked dynamic operational capabilities; it could describe what was stored in the mind, but it could not articulate how that knowledge was actively deployed to execute goals, solve problems, or guide physical interaction.

To overcome HAM’s static limitations, Anderson published Language, Memory, and Thought in 1976, introducing the original ACT architecture. This inaugural formulation augmented propositional networks with a formal production system inspired by Allen Newell’s pioneering work. The original ACT demonstrated that procedural rules could interact directly with an associative semantic network, offering an initial unified account of syntax acquisition, categorization, and logical deduction.

In 1983, Anderson published The Architecture of Cognition, presenting ACT*. This major theoretical update introduced three qualitative forms of knowledge representation: temporal strings, spatial images, and abstract propositions. More crucially, ACT* formalized the mechanisms of proceduralization, composition, and generalization, providing the scientific community with its first end-to-end computational model of skill acquisition—illustrating how deliberate declarative processing transforms systematically into automated procedural performance through sustained practice.

The early 1990s marked a profound philosophical and architectural pivot with the transition to ACT-R (Adaptive Control of Thought-Rational), detailed in Anderson’s 1993 volume Rules of the Mind. Influenced by his work on rational analysis, Anderson discarded arbitrary heuristic search parameters in favor of mathematically optimized Bayesian calculations. ACT-R demonstrated that the human cognitive system is finely tuned to the statistical regularities of its natural ecology; memory retrieval mechanisms and choice utilities reflect rational adaptations designed to maximize task success while minimizing cognitive effort.

Subsequent iterations (ACT-R 5.0, 6.0, and 7.0) shifted the architecture toward a modular, biologically plausible framework aligned with contemporary cognitive neuroscience. Anderson and his collaborators mapped distinct ACT-R components directly onto localized neural substrates (such as associating the retrieval buffer with the hippocampus and ventrolateral prefrontal cortex, the goal buffer with the dorsolateral prefrontal cortex, and the procedural central coordinator with the basal ganglia). This neurofunctional alignment transformed ACT-R into a leading platform for synthesizing behavioral reaction-time data with hemodynamic responses obtained via functional magnetic resonance imaging (fMRI).

6. Theoretical Foundations

Adaptive Control of Thought theory is structurally anchored in the computational theory of mind, information-processing psychology, and rational evolutionary analysis. It explicitly pursues Allen Newell’s celebrated mandate for cognitive science: the construction of “Unified Theories of Cognition.” Newell argued that psychologists should reject fragmented micro-theories of isolated experimental phenomena in favor of overarching computational architectures that explain the entirety of human mental life through a parsimonious set of integrated mechanisms.

A critical pillar of the theory is Herbert Simon’s bounded rationality. ACT assumes that human cognition seeks optimal solutions but is fundamentally constrained by finite processing speed, limited working memory capacity, and incomplete information. Consequently, the architecture’s subsymbolic layer implements continuous heuristic optimizations that balance the expected utility of a cognitive operation against its metabolic and temporal costs.

Epistemologically, ACT resolves the historic conflict between rationalist-symbolic perspectives (which highlight rule-based logical inference) and empiricist-connectionist models (which emphasize distributed associative strength and statistical pattern recognition). By deploying a dual-level architecture, ACT incorporates symbolic structures to handle syntax, compositionality, and hierarchical goal nesting, while using continuous subsymbolic algorithms to capture statistical learning, priming, associative interference, and the gradual decay of unpracticed traces.

Furthermore, the architecture is deeply informed by modern neurobiology. Rather than viewing the brain as an undifferentiated general-purpose computational engine, ACT operationalizes cognition as an orchestrated network of specialized modular subsystems communicating through tightly constrained computational interfaces. The central production system serves as a parallel-matching engine that monitors the state of these peripheral modules and broadcasts commands back to them, mirroring the distributed yet centralized architecture observed in human corticostriatal loops.

7. Key Components, Types & Dimensions

The contemporary ACT-R architecture is structured around specialized modular components, discrete communicative buffers, and multi-tiered computational operations:

• Declarative Module: The long-term repository of factual knowledge. It archives information as discrete symbolic entities termed chunks. Each chunk is defined by a distinct conceptual type and an array of named slots housing specific relational data.
• Procedural Module: The executive engine of cognition. It contains procedural knowledge structured as condition-action productions. It continuously executes 50-millisecond cycle matches against buffer configurations, coordinating internal mental workflows and physical responses.
• Communicative Buffers: Restricted-capacity interfaces through which the procedural module interacts with specialized internal and external modules:
  • Retrieval Buffer: Holds the single declarative chunk currently retrieved from long-term declarative memory.
  • Goal Buffer: Tracks the current mental intention, sub-goal hierarchy, and contextual orientation of the agent.
  • Visual Buffers: Partitioned into the visual-location buffer (representing the spatial coordinates of visual stimuli) and the visual-object buffer (representing the categorical identity and features of an attended stimulus).
  • Manual/Motor Buffers: Manages the planning, preparation, and kinetic execution of motor actions (e.g., keystrokes, mouse clicks, vocalizations).
  • Imaginal Buffer: Maintains dynamic internal representations, working hypotheses, and temporary state-space models during active problem solving.
• Symbolic Layer: The qualitative structural level of the system consisting of explicitly definable chunks and condition-action production rules.
• Subsymbolic Layer: The underlying continuous mathematical algorithms that govern symbolic viability:
  • Base-Level Activation ($B_i$): Quantifies the historical availability of a chunk, calculated as a logarithmic function of its usage history and decay over time.
  • Spreading Activation ($S_i$): Reflects the attentional energy channeled from items currently residing in buffers toward related declarative chunks in long-term storage.
  • Production Utility ($U$): The calculated value determining which production rule fires when multiple rules match the same buffer configuration, based on past payoffs and temporal costs.
  • Noise Parameters: Stochastic perturbations added to activations and utilities to capture human-like behavioral variability, lapses, and occasional performance errors.

8. Examples & Illustrative Cases

The mechanics of Adaptive Control of Thought theory can be clarified by examining how the architecture handles real-world cognitive workflows across developmental and operational stages.

Consider an elementary student learning single-column mental addition, such as calculating 4 + 3. At the initial, declarative stage, the child relies on explicit factual chunks representing basic counting rules (e.g., chunk: start at 4; chunk: add 1 yields 5; counter at 1). The child’s cognitive system executes a series of generic production rules that interpret these chunks sequentially:

1. The goal buffer holds solve 4 + 3.
2. A production fires to retrieve the successor of 4 from declarative memory, loading the chunk 5 into the retrieval buffer.
3. Another production increments an internal counter maintained in the imaginal buffer.
4. This cycle repeats iteratively until the counter reaches 3, at which point the final sum 7 is transferred to the motor buffer to generate a spoken or written response.

This early performance is sluggish, occupies the entirety of the child’s focal working memory, and requires several seconds per calculation. However, as the child solves this exact problem dozens of times, two fundamental shifts occur within the ACT-R architecture. First, a new declarative chunk directly binding the problem elements (4 + 3 = 7) gains high base-level activation through repeated retrieval. Second, through production compilation, the architecture synthesizes the retrieval and motor commands into a single, automated production rule: IF [Goal is to add 4 and 3], THEN [Directly output 7 and pop goal]. At this proceduralized stage, the operation executes within a single 50-millisecond production cycle, completely bypassing intermediate counting steps and freeing working memory for more complex arithmetic reasoning.

Another illustrative case is found in dynamic operational environments, such as an air traffic controller tracking radar blips. The controller’s visual-location buffer detects an unexpected trajectory change on a radar screen, triggering a production that directs the visual-object buffer to shift spatial attention to that locus. Once visual identification occurs, the categorical identity of the aircraft is deposited into the visual buffer. The presence of this identity, coupled with the controller’s active goal buffer (maintain spatial separation between aircraft), activates procedural rules assessing conflict risk. If calculated separation distance violates regulatory thresholds, the procedural module retrieves the emergency vectoring protocol from declarative memory, generates an alternative heading in the imaginal buffer, and immediately drives the vocal motor buffer to issue an urgent radio correction to the pilot.

9. Measurement & Assessment

Assessing the validity and operational parameters of an ACT-R model involves comparing its computational simulations against precise empirical measurements of human performance across multiple methodologies:

Behavioral Latency and Accuracy Profiles: The primary method for testing an ACT model involves recording chronological response latencies, error distributions, and eye-fixation sequences during standardized tasks. Because ACT-R executes operations via quantifiable 50-millisecond production cycles paired with mathematically deterministic retrieval equations, researchers can directly assess the fit between simulated reaction times and human behavioral data. Goodness-of-fit is assessed using coefficients of determination ($R^2$, typically exceeding 0.90 in rigorous models) and Root Mean Square Error (RMSE).

Neurofunctional Alignment via fMRI: Modern ACT-R research interfaces computational models directly with hemodynamic functional neuroimaging data. Specific modules in the architecture have been localized to precise stereotaxic coordinates within the human brain:

• Retrieval Module: Mapped to the left ventrolateral prefrontal cortex (vlPFC) and anterior hippocampus.
• Goal Module: Mapped to the dorsolateral prefrontal cortex (dlPFC).
• Imaginal Module: Mapped to the posterior parietal cortex.
• Procedural Central Executive: Mapped to the basal ganglia (striatum).
• Motor Module: Mapped to the primary motor cortex and cerebellum.

Researchers generate simulated blood-oxygen-level-dependent (BOLD) response curves by convolving the activation time-course of each ACT-R buffer with a standard hemodynamic response function. These predicted curves are then statistically compared against actual neuroimaging scans collected while human participants perform identical cognitive tasks.

Eye-Tracking Alignment: The visual module in ACT-R makes specific predictions regarding visual attention, fixation durations, saccadic trajectories, and perceptual encoding intervals. Researchers test these predictions by comparing model runs against high-resolution eye-tracking recordings, validating the model’s visual-search and scene-parsing mechanisms.

10. Applications & Practical Significance

Adaptive Control of Thought theory has generated practical applications across computer science, human-centered engineering, education, and instructional design.

The most prominent educational application of ACT is the creation of Cognitive Tutors. Developed by Anderson, Kenneth Koedinger, and colleagues, these intelligent tutoring systems monitor students’ problem-solving workflows in domains such as algebra, geometry, and computer programming. By utilizing a technique known as cognitive model tracing, the software runs a internal ACT-R model running parallel to the student’s actions. The tutor identifies the exact production rules—both normative rules and common “buggy” misconceptions—the student is deploying in real time. This allows the system to intervene precisely at the point of cognitive breakdown, providing targeted scaffolding that mirrors human 1-on-1 tutoring. Millions of secondary school students across North America utilize curricula derived from this ACT framework.

In human-computer interaction (HCI) and ergonomics, ACT-R models serve as synthetic users to evaluate digital user interfaces, aircraft cockpits, and automotive dashboards prior to physical prototyping. By executing tasks within a simulated interface, an ACT-R model can predict ergonomic bottlenecks, high-stress visual search sequences, cognitive workload spikes, and potential human error points during system failures.

In defense and aerospace simulation, the architecture is deployed to generate realistic, cognitively valid synthetic agents for military war-gaming and flight simulation. Unlike simple scripted non-player entities, ACT-R-driven agents experience realistic visual constraints, attentional bottlenecks, fatigue, spatial disorientation, and memory decay under stress, providing realistic adversaries and collaborators for human trainees.

11. Research & Empirical Evidence

Empirical support for Adaptive Control of Thought theory has accumulated over several decades of research in experimental psychology and cognitive neuroscience.

In a series of landmark studies on cognitive skill acquisition, Anderson and Mark Singley (1989) investigated the transfer of text-editing skills across diverse computer interfaces. They demonstrated that positive transfer between disparate tasks is governed entirely by the degree of shared production rules between them—providing strong support for the “identical elements” transfer hypothesis originally proposed by Edward Thorndike, reframed in terms of ACT production compilation.

Anderson, Fincham, and Douglass (1997) tracked the chronological neural and behavioral transitions individuals undergo as they acquire complex problem-solving abilities. Their behavioral and neuroimaging data confirmed the precise four-stage transition predicted by ACT theory: from declarative analogical retrieval, to procedural rule synthesis, to independent procedural execution, and ultimately to automated retrieval-driven response generation.

During the 2000s and 2010s, Anderson and his collaborators at Carnegie Mellon published comprehensive fMRI validation studies demonstrating that the temporal engagement of ACT-R buffers accurately predicts hemodynamic activity across distinct brain regions during algebra problem solving, paired-associate memory tests, and navigational tasks. Research by Christian Lebiere and colleagues expanded the framework into decision-making domains, proving that ACT-R’s instance-based learning mechanisms account for human risk preferences and behavioral anomalies in competitive, dynamic game environments.

12. Cultural & Cross-Cultural Considerations

Because Adaptive Control of Thought theory is framed as an architectural specification of the human species’ biological mind, it posits that the fundamental computational mechanisms—such as production cycle speeds, base-level decay rates, spreading activation dynamics, and buffer bandwidths—are human cognitive universals that do not vary across cultures.

However, the content of the declarative memory module and the specific configuration of compiled production repertoires are heavily shaped by cultural, linguistic, and educational environments. Cross-cultural research applying cognitive architectures shows that differences in cultural practices shape the declarative chunks individuals accumulate and the specific procedural heuristics they develop to solve tasks.

For example, comparative mathematical cognition research highlights how the phonological and grammatical structure of counting terms across different languages alters early arithmetic development. In languages with transparent base-10 numerical naming conventions (such as Mandarin Chinese), children develop declarative numerical chunks with lower retrieval latencies and simpler associative pathways than children learning irregular numerical systems (such as English or French). The ACT-R architecture captures these performance differences not by changing its underlying cognitive mechanisms, but by modeling how linguistic differences alter the retrieval dynamics and procedural composition of arithmetic operations.

13. Criticisms, Debates & Limitations

Despite its broad explanatory scope and empirical successes, Adaptive Control of Thought theory faces ongoing criticism and theoretical debate:

• Connectionist Critiques: Proponents of radical connectionism and deep neural network modeling argue that ACT’s reliance on localized, symbolic representations (discrete chunks and production rules) is an outdated abstraction of biological neural processing. They argue that human cognition does not require explicit symbolic rules or discrete communicative buffers; instead, it can be modeled as emergent phenomena within broad, distributed, sub-symbolic neural networks.
• Situated and Embodied Cognition: Theorists from the embodied, situated, and enactive cognitive science paradigms critique ACT for its historical commitment to a centralized, computational processing architecture. They contend that the theory overemphasizes abstract internal manipulations while undervaluing the continuous, dynamic sensorimotor coupling between an organism’s physical body and its immediate environment.
• Parameter Brittleness: Critics note that ACT-R computational models often rely on a substantial number of free parameters (including base-level decay rates, activation thresholds, and noise variances). Skeptics caution that researchers can sometimes adjust these parameters post-hoc to match behavioral data, potentially producing models that fit experimental curves without offering genuine prospective predictive validity.
• Scalability Challenges: While ACT-R excels at modeling circumscribed tasks lasting from seconds to minutes (such as arithmetic, menu selection, or specific flight maneuvers), scaling the architecture to handle lifelong learning, common-sense reasoning, and continuous multi-modal environmental engagement remains an open technical challenge.

14. Related Terms & Distinctions

To contextualize Adaptive Control of Thought theory within the broader landscape of cognitive science, it is helpful to distinguish it from related frameworks:

• Soar: Another leading unified cognitive architecture, developed by Allen Newell, John Laird, and Paul Rosenbloom. While both are production systems, Soar historically rejected declarative-procedural dualism in favor of a single procedural memory structure, used universal subgoaling through decision cycles to resolve impasses, and initially omitted ACT-R’s continuous subsymbolic Bayesian activation mechanisms.
• EPIC (Executive-Process/Interactive-Control): A cognitive architecture developed by David Kieras and David Meyer that emphasizes perceptual and motor execution constraints. Unlike ACT-R, which incorporates a centralized production-cycle bottleneck (firing one production rule at a time), EPIC features an unconstrained, parallel procedural execution system, attributing performance limits solely to peripheral motor and sensory bottlenecks.
• Connectionism / Artificial Neural Networks: Computational paradigms that model cognition through distributed layers of interconnected nodes without symbolic rules or explicit modular buffers. ACT-R combines symbolic elements with subsymbolic activation principles, distinguishing it from pure connectionist architectures.
• Dual-Process Theory (System 1 vs. System 2): A popular psychological framework proposed by Daniel Kahneman and others that divides cognition into fast, intuitive operations (System 1) and slow, deliberative reasoning (System 2). While ACT-R captures similar dynamics (compiled productions execute rapidly like System 1, whereas declarative interpretation requires deliberate steps like System 2), it models both through a single, continuous set of architectural mechanisms rather than two distinct cognitive systems.

15. Summary / Key Takeaways

Adaptive Control of Thought theory stands as one of the most mature, comprehensively validated, and influential cognitive architectures in the history of psychology. By unifying declarative memory structures with procedural production systems—and embedding their interactions within a hybrid symbolic-subsymbolic computational engine—the framework offers an integrated explanation of human cognition across multiple levels of analysis.

From the millisecond-level neural firing of corticostriatal loops to the multi-year progression of students mastering complex academic disciplines, the architecture demonstrates how human intelligence adapts to the statistical structures of its environment. Its successful applications in intelligent tutoring systems, user-interface design, and cognitive neuroimaging underscore its enduring importance as both a theoretical framework and a practical tool in contemporary cognitive science.

References

• Anderson, J. R. (1976). Language, memory, and thought. Lawrence Erlbaum Associates.
• Anderson, J. R. (1983). The architecture of cognition. Harvard University Press.
• Anderson, J. R. (1993). Rules of the mind. Lawrence Erlbaum Associates.
• Anderson, J. R. (2007). How can the human mind occur in the physical universe? Oxford University Press. https://doi.org/10.1093/acprof:oso/9780195324259.001.0001
• Anderson, J. R., Bothell, D., Byrne, M. D., Douglass, S., Lebiere, C., & Quin, Y. (2004). An integrated theory of the mind. Psychological Review, 111(4), 1036–1060. https://doi.org/10.1037/0033-295X.111.4.1036
• Anderson, J. R., & Lebiere, C. (1998). The atomic components of thought. Lawrence Erlbaum Associates.
• Newell, A. (1990). Unified theories of cognition. Harvard University Press.
• Singley, M. K., & Anderson, J. R. (1989). The transfer of cognitive skill. Harvard University Press.