RESEARCH DOSSIER
Cognitive Architectures and Unified Process Models
Abstract. Cognitive architectures seek to explain how memory, perception, learning, cognitive control, and action emerge from reusable computational mechanisms. This fundamental research monograph reconstructs the history of unified cognitive theories and critically compares ACT-R, Soar, EPIC, CLARION, probabilistic and neurally grounded approaches. Primary experiments and original theoretical studies are reviewed across response timing, multitasking, memory retrieval, rule learning, functional neuroimaging, developmental transfer, and neural simulation. The central research question concerns which architecture-level constraints survive changes in task, population, context, and measurement. Evidence supports detailed local process models and several compelling cross-domain integrations, while strong claims of architectural universality require independent generalization, model recovery, transparent training and strategy assumptions, and discriminating experimental contrasts. The synthesis identifies methodological conditions for cumulative cognitive explanation and applications to comparative human–artificial cognition.
Keywords: cognitive architecture; unified theories of cognition; ACT-R; Soar; EPIC; CLARION; production systems; cognitive modeling; executive control; memory retrieval; multitasking; model identifiability; cross-task generalization; computational cognitive neuroscience; human–artificial cognition
Scientific Object and Architectural Commitments
A cognitive architecture is a computationally specified account of the enduring organization that makes diverse cognitive activities possible. Its explanatory commitments concern the kinds of representations a system can maintain, the processes by which those representations change, the resources that constrain simultaneous operations, and the mechanisms connecting perception, memory, deliberation, and action. The research object here is architectural unity across tasks: whether common computational commitments can explain patterns that otherwise require separate accounts of remembering, learning, reasoning, and acting. A working program is essential for making the commitments operational, yet implementation alone supplies no psychological warrant. Newell's program of unified cognition demanded cumulative explanation beyond isolated laboratory phenomena. Newell (1990) Kotseruba & Tsotsos (2020)
An architectural hypothesis is substantive when it prohibits some observations and anticipates others. A fixed production-selection rule, for example, implies competition when several candidate actions are eligible, whereas an architecture permitting multiple simultaneous rule applications can predict a different pattern of overlapping work. The observable record consists of accuracy, response latency distributions, errors, eye and hand movements, interference, adaptation, and sometimes neural measurements. The same record may arise from different internal procedures; researchers must therefore design contrasts that distinguish architectures. The demand for cross-task constraint is central: a parameter or mechanism introduced for a visual-search task should retain an interpretable role when a related theory explains arithmetic or skilled interaction. Anderson et al. (2004) Meyer & Kieras (1997)
This monograph treats ACT-R, Soar, EPIC, CLARION, and selected neurally grounded frameworks as research programs with different explanatory priorities. A unified process model is an executable theory in which a specified set of mechanisms generates linked predictions across more than one cognitive domain. Its standing depends on successful tests of the shared mechanisms, meaningful comparison with alternatives, and transparency about local task assumptions. Historical achievement, engineering utility, biological resemblance, and evidence for human cognition are independently assessed properties. The resulting synthesis attends especially to how modelers move from fitting individual experiments to constructing transportable explanations of cognition in changing environments. Laird et al. (1987) Ritter et al. (2019) Eliasmith et al. (2012)
Historical Formation of Integrated Cognitive Theories
Mid-twentieth-century information-processing psychology formulated cognition as transformations over internal representations. Plans, operations, and feedback became tools for explaining organized activity, as illustrated by Miller, Galanter, and Pribram's analysis of behavior and the subsequent symbolic modeling tradition. The early ambition was to discover reusable procedures for tasks that looked psychologically unlike one another. Contemporary critics and successors inherited this ambition while disagreeing about whether discrete symbols, production systems, parallel activation, or neurally distributed states supply the right explanatory primitives. The historical record shows a succession of proposals concerning stable organization rather than a single discovery that settled cognition's architecture. Miller et al. (1960) Newell (1990) Anderson (1983)
Anderson's Architecture of Cognition consolidated a theory in which knowledge and goal-directed performance were modeled through computationally explicit processes. The 1980s development of Soar by Laird, Newell, and Rosenbloom framed problem solving around operators, impasses, subgoals, and learning, seeking general cognitive competence through a small family of recurring mechanisms. Later revisions enlarged Soar's learning and memory capabilities. These models should be read as evolving scientific programs: an assertion about a 1987 implementation may become false of its 2012 successor. Historical reconstruction therefore identifies the version and empirical setting underlying any claim about what an architecture can represent or learn. Anderson (1983) Laird et al. (1987) Laird (2012)
The 1990s brought detailed tests of architectures against perceptual-motor timing and multitasking. EPIC explicitly represented concurrent peripheral processors and executive coordination, challenging overly simple interpretations of a universal central bottleneck. ACT-R's modular account became increasingly tied to brain-measurement predictions in the following decade. In parallel, CLARION proposed interactions among explicit and implicit knowledge systems. The subsequent literature extended architectural modeling toward neural simulation, structured learning, and hybrid probabilistic methods. These developments changed standards of evidence: explaining a mean latency ceased to be sufficient when trial distributions, transfer, neuroimaging, and implementation constraints could also be tested. Meyer & Kieras (1997) Anderson et al. (2004) Sun & Zhang (2006) Eliasmith et al. (2012)
Levels of Explanation and Architectural Invariants
A unified account requires clarity about its level of description. Marr's distinction among computational problems, representations and algorithms, and physical implementation remains valuable because identical behavior can be generated by multiple algorithms, while a particular algorithm can have multiple physical realizations. Architectural hypotheses typically operate at an intermediate level: they specify allowed state changes and information routes, often informed by cognitive or neural constraints. Cross-level convergence strengthens an explanation when the same model makes discriminating predictions at several levels. Mere correspondence between a module label and a brain region cannot establish an exclusive neural mechanism. Marr (1982/2010) Anderson et al. (2004) Poldrack (2006)
An invariant is a parameter, structural dependency, resource property, or learning operation assumed to remain stable under a stated class of task transformations. It is neither a metaphysical universal nor an assertion that every participant uses the same strategy. In ACT-R, common machinery for production matching and declarative retrieval can be reused while task-specific chunks and rules change. In Soar, problem-space organization and learning procedures are designed to support varied domains. In EPIC, explicit timing and resource assumptions constrain scheduling. The empirical question is whether such invariants survive transfer to tasks, subjects, or contexts that were not used to construct the model. Anderson et al. (2004) Laird (2012) Kieras & Meyer (1997)
Architectural scope must be specified before evaluation. A model may aim to explain practiced adult visual-manual performance while leaving developmental language acquisition outside its initial domain. Expanding the target domain changes what counts as adequate evidence. A compact theory with narrow and reliable predictions can be scientifically preferable to a richly implemented agent whose numerous adjustable mechanisms accommodate almost any sequence. Conversely, a genuinely integrated architecture earns explanatory reach when independent experimental findings follow from shared constraints. This establishes the guiding methodological principle of the monograph: unity is supported by convergent, difficult-to-adjust predictions across heterogeneous tasks rather than by a broad list of demonstrated applications. Kotseruba & Tsotsos (2020) Ritter et al. (2019) Newell (1990)
Architecture, Process Models, and Task-Specific Programs
An architecture supplies the computational grammar of possible models; a task model supplies the knowledge, representations, goals, and strategies that instantiate that grammar for a particular experiment. A modeler may hold core retrieval dynamics fixed while varying the number of practiced associations or the task's stimulus mapping. This distinction helps prevent a favorable fit from being credited entirely to the architecture when task-specific strategies do much of the explanatory work. A fair evaluation reports which components were fixed, estimated, or programmed, and what behavioral signatures would be impossible under alternative architectural specifications. Anderson et al. (2004) Ritter et al. (2019) Kotseruba & Tsotsos (2020)
Models are sometimes classified as descriptive, generative, normative, or mechanistic. Architectural theories usually aspire to generative and mechanistic roles: they specify a sequence of operations whose execution yields predicted actions and timings. A model's normative role, such as optimizing reward, is an additional commitment that can be compared with empirical adaptation. Observational equivalence presents a persistent difficulty. Different procedures can agree on the same average reaction time while implying distinct error patterns, learning curves, or responses to interruption. Researchers should construct tasks that amplify the differences, then test on independent observations rather than selecting the best post hoc explanation. Collins & Frank (2013) Salvucci & Taatgen (2008) Newell (1990)
The architecture–task separation also clarifies comparisons between symbolic cognitive theories and machine-learning systems. An artificial network trained on many examples might solve a task effectively without instantiating the same latent computations that human participants use. Conversely, a psychologically precise production system may predict human errors while performing poorly as an unconstrained engineering agent. Claims about cognitive similarity require common tasks, matched information exposure, explicit learning histories, and analysis of which internal mechanisms account for observed behavior. Technical success and psychological explanation can coexist, but their evidential criteria must be independently articulated. Tenenbaum et al. (2011) Eliasmith et al. (2012) Kotseruba & Tsotsos (2020)
Knowledge Representations, Productions, and Control
Production-system architectures represent procedural knowledge through condition–action structures that detect relevant states and initiate transitions. A rule's explanatory role depends on its activation conditions, conflict resolution, timing, and access to memory. Such models can express sequences of goal-directed behavior while reusing a shared interpreter across tasks. They also raise an underdetermination problem: a researcher can often encode an experimental strategy in several alternative sets of productions. Strong tests therefore examine architectural constraints, such as competition between candidate operations and retrieval costs, rather than celebrating the model's ability to reproduce a familiar action sequence. Anderson (1983) Anderson et al. (2004) Laird et al. (1987)
Executive control concerns how a system prioritizes, initiates, monitors, and switches operations under changing task conditions. Production-based accounts can distribute these functions across goal representations and rule-selection machinery, while other architectures introduce richer scheduling strategies or explicit monitoring circuits. The conflict-monitoring model of Botvinick and colleagues made the relation between competition and subsequent control adjustment computationally testable, illustrating how an executive explanation can generate predictions for behavior and neural activity. Its empirical support is evidence for specified control mechanisms, not proof that an entire architecture has been uniquely identified. Botvinick et al. (2001) Meyer & Kieras (1997) Cooper & Shallice (2000)
The organization of control is particularly visible in sequences of actions that must survive distraction or interruption. Cooper and Shallice simulated routine action selection with hierarchically organized competing schemas and explored impairments after simulated lesions. Their work links action disorganization to an executable causal hypothesis about selection dynamics. More generally, architectural claims about control should predict which errors become likely when an operation is delayed, a goal loses activation, a competing action gains support, or supervisory control is burdened. These counterfactual predictions give the architecture testable content that a purely verbal account of central executive capacity lacks. Cooper & Shallice (2000) Salvucci & Taatgen (2008) Botvinick et al. (2001)
ACT-R: Modular Cognition and Production Selection
ACT-R formalizes cognition as interactions among specialized modules whose current contents become available through bounded buffers. In the influential integrated account by Anderson and colleagues, visual, motor, goal, and declarative-memory components communicate with a production system that selects actions using accessible buffer information. The resulting simulations can predict sequences of cognitive states and observable response times. Modules are theoretically useful because they impose restrictions on the information a process can immediately use; without such restrictions, the theory could construct convenient instantaneous dependencies between any task components. ACT-R's distinctiveness is therefore in the combination of representations, retrieval, production selection, and temporal constraints rather than in the bare claim that cognition is modular. Anderson et al. (2004) Ritter et al. (2019)
The 2004 integrated theory connected model components with cortical activation patterns, offering convergent evidence beyond behavior. Such mappings are hypotheses about component involvement, not one-to-one anatomical identities. The correspondence between a buffer's modeled activity and a measured hemodynamic response involves assumptions about measurement, temporal convolution, and task decomposition. Researchers must ask whether neural predictions were specified before observing the relevant data and whether alternative process models would yield comparable activation. The architecture's strongest claims concern combined restrictions on behavior and neural timing when one theory is evaluated across several tasks. Anderson et al. (2004) Poldrack (2006)
ACT-R also demonstrates why integration entails tradeoffs. Specialized modules capture differences between perceptual and memory processing, while a common procedural mechanism organizes their interaction. Increasing modular detail can improve empirical fidelity yet introduce extra discretionary modeling choices. Careful evaluation separates parameters defined by the architecture from quantities fit for each subject and assumptions supplied by the modeler. A genuinely cumulative test should reuse activation equations and operation timing across tasks, reveal failures as well as favorable outcomes, and establish whether improvements to one domain preserve prior results. The review by Ritter and colleagues documents ACT-R's continuing use in basic and applied modeling and its explicit computational commitments. Ritter et al. (2019) Anderson et al. (2004)
Declarative Memory, Activation, and Retrieval Dynamics
Memory is an architectural test bed because recall depends simultaneously on stored experience, retrieval cues, interference, and current goals. ACT-R-style retrieval accounts represent knowledge in chunks and model accessibility through activation that depends on prior exposures and contextual relations. A memory operation is therefore neither an unconstrained lookup nor a direct translation of a psychological label into a software variable. Its predictions concern latency, failure, and competition among candidates as histories and cues change. Researchers can test the same retrieval assumptions in arithmetic, recognition, skill performance, and instruction following, where a shared memory mechanism ought to produce related signatures. Anderson et al. (2004) Anderson (1983)
Working-memory limitations provide an especially demanding comparison because many different mechanisms can reduce simultaneous task performance. A limited activation resource, interference among representations, restricted time for refreshing, and competition for processing resources can generate overlapping empirical patterns. Daily, Lovett, and Reder examined a source-activation account of individual differences in working-memory performance; such work shows that claims about a capacity limitation require specific model-based interpretation. A test that changes memory load and competing operations can discriminate among candidate resource assumptions better than a simple correlation between span and a second task. Daily et al. (2001) Baddeley (2000)
The integration problem arises when retrieval is modeled within an active task. Which representation becomes accessible affects production selection; the selected action alters the context of subsequent retrieval. Feedback loops of this kind make it possible for one architecture to produce nonlinear learning and variable error sequences, but they also complicate parameter identification. A model that explains a short retention interval through high activation might explain it equally through changed cue associations. Longitudinal within-person data and experimental manipulation of interference are therefore valuable. A useful unified account predicts not merely whether memory succeeds, but the consequences of memory failures for control, response timing, and future learning. Anderson et al. (2004) Daily et al. (2001) Salvucci & Taatgen (2008)
Soar: Problem Spaces, Impasses, and Learning
Soar addresses the ambition of general cognitive competence through an organized decision cycle over representations of states, operators, and goals. When available knowledge does not determine a next step, an impasse can initiate a substate in which the system reasons about what is missing. Resolution can produce knowledge that changes later decisions. Laird, Newell, and Rosenbloom presented Soar as an implemented proposal for a broad architecture rather than as a proof that human cognitive organization had been discovered. The architecture's explicit organization is valuable because researchers can inspect the intermediate decisions that generate its behavior and ask which assumptions cause its successes and errors. Laird et al. (1987) Rosenbloom et al. (1991)
Learning in the Soar tradition includes the chunking of results from reasoning episodes into reusable procedural knowledge, while later versions also include multiple memory and learning systems. The mechanism promises transfer when future problems reencounter structurally related conditions. Such transfer is empirically meaningful only when the relation between old and new problems is specified independently, and when knowledge gained during training is separated from hand-coded domain competence. Comparisons with human skill acquisition should consider improvement trajectories, negative transfer, error recovery, and the cost of constructing new knowledge. A demonstration of an agent solving many tasks shows architectural capability; it is stronger psychological evidence when task behavior and learning processes match humans under comparable experience. Laird (2012) Rosenbloom et al. (1991)
Soar's historical evolution illustrates the problem of architectural versioning. A source about its early universal-subgoaling proposal cannot be generalized without qualification to later memory, reinforcement-learning, and perceptual extensions. The comprehensive 2012 treatment documents the expanded architecture as a changing research object. This matters to reproducibility: a claimed successful replication of a Soar account requires the program version, task knowledge, learning state, environmental interface, and evaluation code. It also matters for theoretical evaluation, because adding components can extend empirical scope while changing the simplicity and explanatory costs of the original proposal. The proper unit of evidence is a specified model under specified constraints, situated within its architectural lineage. Laird (2012) Laird et al. (1987) Kotseruba & Tsotsos (2020)
EPIC: Concurrent Processing and Perceptual-Motor Constraints
EPIC, the executive-process interactive control architecture, emphasizes the detailed temporal organization of perceptual, cognitive, and motor operations. Meyer and Kieras proposed multiple peripheral processors operating under executive scheduling, with production-rule cognition able to support forms of concurrent procedural activity. The architecture's value lies in its explicit assumptions about what may overlap, what must wait, and how tasks coordinate motor output. These commitments produce observable predictions measured in milliseconds and in the ordering of responses. The account challenges the adequacy of a single immutable central bottleneck as a universal interpretation of interference, while retaining the importance of task-specific sensory and motor constraints. Meyer & Kieras (1997) Kieras & Meyer (1997)
The psychological refractory-period paradigm tests performance when responses to closely spaced stimuli must be coordinated. Different temporal accounts can predict similar increases in the second response time; the discriminating issue concerns how experimental manipulations alter the overlap of processing stages. EPIC modeling investigates how strategic scheduling, practice, and response priorities shape those patterns. A precise model must specify both architecture and participant strategy, since a participant can choose a cautious sequential response policy even when parallel operations are structurally possible. This is a general warning against inferring a fixed cognitive mechanism directly from a latency difference. Meyer & Kieras (1997) Schumacher et al. (2001)
Evidence about near-perfect dual-task performance under practiced and favorable circumstances supports conditional flexibility, not unlimited general parallelism. Schumacher and colleagues reported experiments in which some participants approached time sharing in particular basic choice tasks after practice. Their result imposes an important boundary condition on theories that assume unavoidable selection interference under every situation. Yet the tasks' response mappings, training, and participant characteristics constrain interpretation. A unified architecture should explain both high interference and its reduction under specified conditions, preferably without replacing one fixed bottleneck rule with unlimited arbitrary scheduling freedoms. EPIC's methodological contribution is an explicit temporal architecture that permits such competing predictions to be tested. Schumacher et al. (2001) Meyer & Kieras (1997) Kieras & Meyer (1997)
CLARION: Explicit and Implicit Computational Processes
CLARION approaches integration through interactions between explicit symbolic and implicit distributed forms of knowledge. Sun and Zhang used this architecture to model varieties of everyday reasoning, including rule-based and similarity-based components. Its distinct question concerns how accessible rule knowledge and more gradual associative processes cooperate to support behavior. This is an architectural hypothesis about representations and pathways of learning, rather than merely a division of tasks into conscious and automatic categories. A model becomes testable when different amounts of instruction, practice, similarity, and feedback predict systematic shifts in the relative contribution of these mechanisms. Sun & Zhang (2006) Kotseruba & Tsotsos (2020)
A dual-representation account can explain why people sometimes articulate a rule yet respond according to accumulated experience, or apply a learned association without being able to report its basis. However, the inference from verbal report to internal representational type is indirect. Performance differences may reflect task demands, confidence, response criterion, or language rather than a pure partition of cognitive processing. Computational integration should make distinct predictions concerning learning rate, generalization, and rule switching under interventions. The strongest evidence would compare variants with or without explicit–implicit interactions using independent tasks while controlling model flexibility. Sun & Zhang (2006) Taatgen & Anderson (2002)
The broader comparative significance of CLARION is that explanatory unity need not require one representational format. Heterogeneous architectures can still impose cross-task constraints if the interfaces and learning mechanisms are stable and operationally specified. The relevant comparison with ACT-R and Soar is therefore mechanistic: which observed transitions depend on rule competition, associative activation, or interaction between representations? Direct architectural contests remain difficult because they differ in task encodings and available optimization. A rigorous program would predefine common benchmark environments and training histories, then record predictive accuracy and the amount of task-specific programming required for each architecture. Sun & Zhang (2006) Ritter et al. (2019) Laird (2012)
Probabilistic Inference and Structured Abstraction
Probabilistic cognitive models describe how an observer could infer hidden structure from uncertain evidence and revise expectations after additional observations. Tenenbaum and colleagues emphasized the combination of statistical learning and structured representations in concept acquisition, language, and causal understanding. These accounts broaden the architectural question by asking which abstractions can be learned and represented, and how uncertainty influences their reuse. A Bayesian computation specifies a relationship between hypotheses, evidence, and prior assumptions; it does not by itself identify the cognitive algorithm or neural mechanism that performs the calculation. The architecture must add process commitments about representation, approximation, memory, and computational cost. Tenenbaum et al. (2011) Marr (1982/2010)
Formal unification is possible when one inference principle explains learning and performance across domains, but similar equations can conceal distinct psychological processes. A model fitted to causal judgments, for example, may presuppose a structured hypothesis space that is different from the one required for motor decisions. The source of that structure and the time needed to search it are part of the explanatory burden. The distinction between a computational-level ideal and a process-level implementation matters particularly when reaction times, errors, developmental changes, and strategy switching must also be predicted. An architecture that invokes probabilistic inference should state how those operations interact with memory and control during actual tasks. Tenenbaum et al. (2011) Collins & Frank (2013)
Cross-framework comparison can test whether probabilistic structure and production-based control play complementary roles. Collins and Frank developed a context–task-set account and a neurally explicit model to examine how learned structures can guide control and generalization. Their combined experimental and simulation approach exemplifies a constructive bridge between levels rather than a claim that one algorithm suffices for all cognition. The important methodological demand is to evaluate how learned task sets transfer to untrained situations and whether abstract structure improves predictions beyond a model that memorizes local stimulus–response associations. Such tests distinguish flexible abstraction from retrospectively convenient labels assigned to familiar results. Collins & Frank (2013) Frank & Badre (2012)
Neural Constraints and Mechanistic Correspondence
A cognitive architecture can be compatible with neural data at several degrees of strength. A weak account assigns components to broad cortical networks after a behavioral model is constructed. A stronger account specifies temporal and anatomical predictions from independent operations, then evaluates them against brain measurements alongside responses. Anderson and colleagues associated ACT-R modules with cortical regions in an integrated theory, thereby taking on commitments that could be contradicted by neuroimaging. Interpretation nonetheless requires attention to the indirect relation between model activity, neural computation, and recorded signal. The same measured region may support diverse operations, making simple reverse inference unreliable. Anderson et al. (2004) Poldrack (2006)
Hierarchical control theories provide a second route from abstract processes to neural circuitry. Frank and Badre modeled interactions of prefrontal representations and basal ganglia gating mechanisms; their companion investigation tested predictions using functional magnetic resonance imaging. The theory connects learning about conditional rules with a proposed implementation of gating and hierarchical selection. Its value for architecture research lies in jointly specifying an abstract computational problem and possible neural mechanisms. However, correspondence between neural activity and a candidate process still leaves open alternative models sharing the same network response. Independent interventions, recovery analyses, and temporally informative measurements can further discriminate among explanations. Frank & Badre (2012) Badre & Frank (2012)
Biological plausibility should be evaluated as a graded explanatory property. A symbolic architecture can achieve accurate behavioral timing without modeling synaptic learning, and a large neural simulation can reproduce several tasks without capturing human developmental history. These are different successes with different scientific limitations. Comparisons should therefore state the biological scale being modeled, the measurement used to validate it, the extent of parameter calibration, and the independent evidence constraining each mapping. An architecture is strongest when its cross-level commitments limit what it can explain after seeing the outcome, instead of providing a flexible vocabulary for redescribing any observed brain region. Eliasmith et al. (2012) Marr (1982/2010) Poldrack (2006)
Large-Scale Neural Simulation and the Spaun Program
Eliasmith and colleagues introduced Spaun as a large-scale neural model performing multiple tasks through a common visual input and motor output arrangement. The original Science report describes a model of approximately 2.5 million neurons demonstrating eight tasks, a noteworthy attempt to integrate neural implementation and cognitive capability within one working system. These figures describe the reported model and its experimental demonstrations, not the number or organization of neurons required by a biological person. The evaluation is an existence demonstration that selected mechanisms can support diverse behaviors under the implemented conditions. Its direct psychological support depends on the precision of human comparisons and the constraints maintained across tasks. Eliasmith et al. (2012)
The neural engineering approach exposes organizational choices that remain implicit in higher-level symbolic models: how information passes among functional components, how population representations are transformed, and how motor activity emerges from dynamic neural states. The advantage is explicit multi-level simulation. The challenge is that high-dimensional neural systems can still be adjusted through learned weights, selected representations, and custom task scaffolds. A fair comparison with ACT-R or Soar therefore asks which properties are fixed by the architecture, which by individual models, and which by external interfaces. Without that separation, the richness of the implementation can obscure rather than settle explanatory competition. Eliasmith et al. (2012) Kotseruba & Tsotsos (2020)
Large neural models also foreground the difference between reproducing a task and explaining the distinctive pattern of human failure. Human participants show variability, strategy differences, fatigue, imperfect memory, and context-dependent interference. A simulator that completes eight tasks does not automatically predict those distributions. Future tests can place the model in experimentally matched conditions, measure full response trajectories, remove components, and compare effects with human perturbation or lesion evidence. Such converging tests would clarify whether the architecture's apparent breadth reflects common cognitive organization or a sequence of independent engineered capabilities sharing one platform. The distinction is central to the scientifically responsible comparison of artificial agents with human cognition. Eliasmith et al. (2012) Cooper & Shallice (2000) Newell (1990)
Executive Control, Conflict Monitoring, and Task Sets
Executive control allows learned operations to be organized according to task goals and changes in the environment. Architectures differ in whether they treat control as a dedicated supervisory mechanism, a distributed product of competition, a procedure of resource allocation, or an emergent property of learned task sets. Botvinick and colleagues proposed a conflict-monitoring account linking competition to adjustments in control, using computational simulations with behavioral and neural implications. The broader lesson is that an executive label must be decomposed into detection, updating, selection, and implementation operations, each with measurable consequences. A unified model should account for the relation among these operations rather than naming their collective outcome. Botvinick et al. (2001) Cooper & Shallice (2000)
Structured task-set learning offers a route to flexible control that avoids inventing a new decision rule for every situation. Collins and Frank modeled the discovery and reuse of latent context–task representations and reported experiments bearing on transfer and choice sequences. Their account demonstrates that the history of learning can change which operations become available during later tasks. The diagnostic observations are therefore both immediate decisions and longitudinal changes after structured practice. When a participant transfers appropriately to a new context, an architecture should identify which representation is reused, how the context is recognized, and what cost is incurred when the assignment is wrong. Collins & Frank (2013) Frank & Badre (2012)
Competing accounts of control can be experimentally separated by manipulating conflict frequency, rule-switch demands, feedback structure, and interference timing. A model that predicts improved selection after conflict must specify the timescale of adjustment and potential competing effects of practice, expectancy, or stimulus repetition. The same architectural mechanism should be evaluated in several tasks with common parameter constraints. The principal scientific risk is treating each observed sequential effect as evidence for control after adding an adjustable update rule. Strong explanatory success requires out-of-sample predictions and a clear separation between theoretical variables and convenient measures. Botvinick et al. (2001) Salvucci & Taatgen (2008) Ritter et al. (2019)
Learning, Development, and Transfer Across Domains
Learning is central to an architecture's claim of unity because the same system must organize unfamiliar behavior and alter it through experience. An architecture may learn production utilities, change retrieval accessibility, compile practiced sequences, update associations, or infer a latent task structure. These mechanisms can generate similar improvements in speed but different error paths and transfer patterns. Researchers should therefore analyze individual learning curves and sequences of response strategies, not solely pretraining and posttraining averages. The essential question is what changes within a stable architecture as a result of experience and whether that change has measurable consequences on tasks beyond those used for training. Anderson (1983) Laird (2012) Collins & Frank (2013)
Language development provides an instructive cross-domain test. Taatgen and Anderson presented a hybrid ACT-R model of English past-tense learning that reproduced aspects of a U-shaped developmental pattern without relying on direct corrective feedback or certain previously assumed vocabulary changes. The model is a computational explanation of a particular developmental pattern; the study does not demonstrate that children's full linguistic development is generated by ACT-R. Its scientific importance lies in specifying how a limited family of learning and memory operations might account for nonmonotonic performance. Replication should assess input assumptions, individual differences, and alternative learning mechanisms rather than merely reproducing the qualitative curve. Taatgen & Anderson (2002)
Developmental and cross-cultural extensions raise demanding generalization questions. A learning architecture tuned to adults using familiar symbols may need substantially different perceptual interfaces, reward histories, and linguistic knowledge to predict children's or multilingual participants' behavior. Such changes may be legitimate features of the environment, but they should not silently become new architecture-specific exceptions. Transfer studies can separate structural continuity from training-history effects by holding a common learning procedure fixed while varying contexts and available experiences. Strong evidence would include successful prospective predictions for new participants and tasks, systematic failures where the model expects interference, and a record of which predicted developmental transitions actually occurred. Taatgen & Anderson (2002) Tenenbaum et al. (2011) Kotseruba & Tsotsos (2020)
Response-Time Distributions and Observable Process Signatures
Response time is one of the most informative dependent variables in architecture research because its variability contains information about multiple processing stages. Averaged latency can conceal mixtures of retrieval success, strategic waiting, response caution, and occasional slow operations. A process model must specify how trials generate timing observations and what distributional shape results. Error rates, response order, and speed–accuracy conditions further constrain interpretation. EPIC and ACT-R supply explicit temporal mechanisms, but task programs still determine many predicted trajectories. Testing should compare full observed distributions when feasible and distinguish latency explained by fixed architecture from latency explained by participant-specific strategy parameters. Meyer & Kieras (1997) Anderson et al. (2004)
Distributional evaluation can expose architectural assumptions that fit means while failing in tails or conditional patterns. For instance, a model with a single constant processing delay may reproduce an average interference cost, yet fail to describe how that cost changes when response priorities or memory demands vary. Researchers can examine quantile functions, conditional accuracy, and sequential dependencies in preregistered or held-out data. Each added parameter should receive a clear psychological interpretation and be subjected to identifiability checks. What matters is not numerical complexity alone, but the way modeled operations account for systematic patterns that follow experimental manipulation. Kieras & Meyer (1997) Salvucci & Taatgen (2008)
Temporal signatures must also be interpreted with measurement realities in view. Display timing, input devices, response collection, practice blocks, and participant exclusion rules can materially change the observed distribution. Computational experiments are reproducible only when the stimulus schedule, model version, timing conventions, random seeds, and analysis scripts are available. Comparison across laboratories additionally requires matching task definitions while preserving potentially informative differences in apparatus and population. A unified architecture should predict which patterns are robust to such changes and where it expects altered sensory or motor constraints to shift the result. Kotseruba & Tsotsos (2020) Ritter et al. (2019) Schumacher et al. (2001)
Multitasking and the Central-Bottleneck Debate
Multitasking research offers an unusually sharp test because architecture theories differ on whether response selection must proceed sequentially. The psychological refractory-period tradition examines responses to temporally overlapping stimuli, while EPIC emphasizes the role of adaptable coordination and peripheral constraints. The existence of dual-task interference does not by itself locate a fixed central stage; additional experiments must identify what kind of overlapping activity was possible and how participants scheduled it. The experimentally relevant contrast is between specific resource organizations, not the general proposition that two tasks are harder than one. Meyer & Kieras (1997) Kieras & Meyer (1997)
Schumacher and colleagues demonstrated that some practiced participants could show near-perfect time sharing for particular simple choice tasks, with instructions and response priorities influencing interference. This result limits a universal reading of an immutable selection bottleneck; it does not imply that arbitrary mental tasks can always proceed without competition. Architecture comparisons must represent the effects of practice, stimulus compatibility, task preparation, and manual or vocal response requirements. A successful theory should jointly explain persistent interference in one condition and its reduction in another while using consistent assumptions about which operations share resources. Schumacher et al. (2001) Meyer & Kieras (1997)
Salvucci and Taatgen's threaded cognition model offered another explicit integration account, proposing concurrent task threads coordinated through a serial procedural resource and other available processors. The model can generate interference patterns and applications extending from laboratory dual-choice tasks toward driving and distraction. Its value is the production of detailed predictions about when processing streams collide and when they can be interleaved. Its limitations depend on whether the specified threads and resources anticipate new behavioral records without discretionary reprogramming. The broader debate is resolved locally through experimentally discriminating predictions rather than by choosing parallelism or seriality as a universal slogan. Salvucci & Taatgen (2008) Schumacher et al. (2001)
Neuroimaging, Lesion Models, and Causal Intervention
Neural measures can strengthen tests of cognitive architecture when a model predicts more than a generic increase in regional activation. ACT-R's module-related hypotheses and corticostriatal models of hierarchical control illustrate two approaches to mapping computational operations to brain signals. Investigators should specify the mapping function, the spatial and temporal resolution of the measure, and the experimental manipulation that could distinguish proposed mechanisms. Poldrack's analysis of reverse inference identifies the danger of treating activation in a region as proof of a uniquely associated process. A stronger inference tests the joint pattern of predicted behavior and measurement under conditions where competing models disagree. Anderson et al. (2004) Poldrack (2006) Badre & Frank (2012)
Perturbation evidence can move beyond correlation by asking how a system changes after a targeted component is interrupted or altered. Cooper and Shallice's computational lesioning of routine action-control mechanisms is valuable because it generates qualitative predictions about action errors. In human studies, lesion interpretation still involves heterogeneity, plasticity, and functional reorganization. A simulated impairment matches an observed syndrome most convincingly when the model predicts the affected sequence, preserved functions, and recovery limits prospectively. Researchers should avoid assigning an entire cognitive process to a single component solely because manipulating that component disrupts one task. Cooper & Shallice (2000) Marr (1982/2010)
For biologically elaborate simulations, causal tests also require an explicit mapping between simulated and measured units. Removing a software module is an intervention on an implemented theory, but its relation to a physiological lesion may be indirect. A good evaluation includes controls for total computational resource loss and comparison with alternative perturbations. When several models survive the same behavioral and neural tests, the evidence supports a class of mechanisms rather than a uniquely identified architecture. This calibrated interpretation permits genuine progress without replacing measurement uncertainty with unjustified structural confidence. Eliasmith et al. (2012) Frank & Badre (2012) Poldrack (2006)
Individual Differences and Capacity Assumptions
Cognitive architectures frequently begin with an idealized participant and a common set of parameters. Psychological populations, however, differ in practice history, sensory precision, response policy, retrieval accessibility, and strategy choice. A unified theory must determine which differences reflect architectural variation and which can arise from identical mechanisms operating over different experience. Daily, Lovett, and Reder's source-activation modeling illustrates one route to explaining working-memory performance differences by changing interpretable resource characteristics. Such accounts make predictions about how an individual responds to load and interference, which are more informative than a generic trait label. Daily et al. (2001) Anderson et al. (2004)
Between-person effects require designs that distinguish genuine parameter variation from poor estimation. A participant with noisy responses may appear to have a slow retrieval parameter because the model lacks an error process, while a shift in response caution may be attributed to limited capacity. Hierarchical models can stabilize estimation, but pooling creates its own assumptions about population structure. Researchers should assess parameter recovery on simulated data, compare within-person repeated sessions, and report uncertainty rather than identifying latent traits from a single model fit. The explanatory target is the joint distribution of behavioral variability and measured task demands. Daily et al. (2001) Collins & Frank (2013)
Cross-population generalization is critical for scientific claims of universality. Tasks derived from university laboratories may presuppose familiarity with digital interfaces, linguistic instructions, or timed testing. Differences across age groups or cultural settings can therefore reflect measurement conditions alongside mechanisms. A transportable architecture should state which parameters and representations may legitimately vary while retaining invariant process predictions. Evidence from new populations should be used to test those limits rather than to redefine them after observing the data. This research strategy connects cognitive architecture with construct validity, differential psychology, and longitudinal assessment without treating any one score as an architectural fingerprint. Kotseruba & Tsotsos (2020) Newell (1990) Ritter et al. (2019)
Identifiability, Model Recovery, and Predictive Validation
A computational architecture may produce convincing behavior while remaining observationally indistinguishable from several competitors. Identifiability concerns whether particular mechanisms or parameters can be recovered from the available observations under the assumed model. The problem is severe when architectures offer many task-specific representations and strategic decisions. A good-fitting response-time distribution may be generated by delayed retrieval, increased decision caution, or changed scheduling. Researchers should first test synthetic recovery: can an analysis correctly identify models and parameter values from data generated under known alternatives? Failure implies that the real experiment cannot settle the intended architectural question. Kotseruba & Tsotsos (2020) Anderson et al. (2004)
Predictive validation should use tasks or conditions withheld from fitting, ideally designed before inspecting the test results. Cross-validation within the same experiment is useful for some purposes, while cross-domain prediction is particularly demanding for an architecture claiming unification. Model comparison must account for the freedom to alter task knowledge, add production rules, or refit activation parameters after each failure. Complexity is not merely a count of numerical parameters; discretionary coding and unreported strategy search also expand the effective hypothesis space. A convincing result identifies which core restrictions were preserved and why the competing architecture could not explain the observation equally well under its own accepted assumptions. Newell (1990) Ritter et al. (2019) Laird (2012)
The same discipline applies to comparisons involving neurally grounded and probabilistic architectures. A neural model may incorporate extensive pretraining and task-specific transformations, whereas a symbolic model may receive explicit instructions and hand-coded representations. Such comparisons are interpretable only if the asymmetry in prior knowledge and adaptation opportunity is disclosed. Researchers can report both predictive performance and the engineering effort or supervision required to obtain it. The appropriate scientific conclusion follows the evidence actually collected: success on a shared benchmark establishes relative performance under its conditions, while a mechanism-level explanation requires discriminating tests that connect model internals with psychological measurements. Eliasmith et al. (2012) Tenenbaum et al. (2011) Kotseruba & Tsotsos (2020)
Cross-Task Generalization and Architecture Benchmarks
A general architecture must be judged by its ability to explain relationships among tasks. An extensive demonstration portfolio is informative, but each task can still be supported by separately constructed knowledge and interface code. Cross-task evaluation should specify what remains fixed: learning rules, resource mechanisms, retrieval parameters, representational formats, or timing assumptions. It should also specify what changes because the task itself requires different information. A benchmark that tests only endpoint accuracy misses the architecture's intended psychological content. Response trajectories, interference, learning curves, and error structures are often more discriminating. The survey by Kotseruba and Tsotsos helps situate architectural diversity and practical applications within this broader evaluative problem. Kotseruba & Tsotsos (2020) Newell (1990)
One useful research design trains models and participants on a family of tasks, introduces an untrained manipulation, and compares model predictions made before observing new outcomes. The manipulation might alter simultaneous response demands, associative interference, or the availability of a learned task structure. Architecture-level generalization requires more than producing a compatible answer; it requires preserving the machinery that explains earlier domains while correctly anticipating the new domain's characteristic errors and time course. Transparent reporting of model construction time and task-specific edits helps distinguish transfer from post hoc tailoring. When an architecture fails, recording that failure can guide theory revision and preserve cumulative constraints. Salvucci & Taatgen (2008) Collins & Frank (2013) Ritter et al. (2019)
Benchmark diversity should also account for the ways different architectures encode tasks. A human-like perceptual interface may favor models with explicit sensors and motors, whereas a purely symbolic puzzle may favor rule-based systems. Comparison can therefore include matched observation and action protocols while reporting differences in preexisting representation and training. The strongest conclusions remain conditional on which cognitive capacities the benchmark samples. No finite battery establishes general intelligence or a complete model of mind, but a well-designed battery can rule out specific constraints or reveal shared mechanisms that survive several independent experimental conditions. Laird (2012) Eliasmith et al. (2012) Kotseruba & Tsotsos (2020)
Ecological Cognition, Skilled Interaction, and Human Factors
Architecture studies began with tightly controlled laboratory tasks but increasingly consider performance in environments where goals, displays, and interruptions change continuously. Applied models must represent the interface between perception, movement, external information, and task control. EPIC's emphasis on perceptual-motor timing is particularly relevant to human–computer interaction, and threaded cognition has been applied to driver distraction and divided attention. The practical contribution is a way to derive consequences of interface design from specified process limits. Ecological prediction, however, requires validation in the relevant setting: a model calibrated on a simple choice experiment may need further evidence before its predictions are used in safety-critical activities. Kieras & Meyer (1997) Salvucci & Taatgen (2008)
Situated tasks also expose assumptions about what belongs inside the modeled cognitive system. Written reminders, interface cues, and automated suggestions can reduce memory or control demands; their effects depend on how a participant perceives and uses them. An architecture may legitimately treat those supports as environmental inputs, but it must specify the interaction rather than crediting unexplained performance gains to internal cognition. Research designs should vary cue availability, interface reliability, workload, and learning history while monitoring actual changes in behavior. The resulting tests can reveal which mechanisms are robust across laboratory and naturalistic contexts. Marr (1982/2010) Ritter et al. (2019) Kotseruba & Tsotsos (2020)
A coherent applied research program distinguishes descriptive simulation from validated predictive engineering. A program may generate plausible eye movements and keystrokes but fail in uncommon interruptions or among novice users. Evaluation should compare event timing, error categories, adaptation after exposure, and generalization to altered interfaces. Human-factors models gain credibility when they identify which interface change should improve performance and when that prediction is borne out prospectively. This approach also connects cognitive architecture to the practical study of collaboration with artificial tools while retaining the requirement for independent evidence of human cognitive processes. Kieras & Meyer (1997) Schumacher et al. (2001) Salvucci & Taatgen (2008)
Artificial Systems and the Comparative Psychology of Architecture
Artificial systems provide implemented examples of how shared computational resources can support diverse behavior. Classical symbolic systems, neural simulations, and contemporary learned models permit researchers to manipulate representations, learning histories, and resource constraints in ways that are difficult in human experiments. These interventions make architectural theories unusually transparent at the level of program execution. The central comparative question is functional and causal: which computational commitments are shared, which behaviors arise through different mechanisms, and which human predictions follow from the comparison? A successful artificial performance does not establish that the human brain uses its algorithm, yet a controlled artificial model can test whether a proposed mechanism is sufficient to generate an observed pattern. Laird (2012) Eliasmith et al. (2012) Tenenbaum et al. (2011)
The historical horizon of the Artificial Era expands the relevance of cross-system architecture studies. Human–artificial cognitive partnership requires models of joint task allocation, communication timing, delegation, error correction, and reliance on external memory. These are empirically tractable questions when the participating systems and interaction protocols can be specified. An architectural model might predict when human attention is displaced by an automated suggestion or when an artificial partner's response latency changes human scheduling. The inference should follow observed interaction and transparent implementation, with explicit testing of alternatives such as simple cueing or task familiarity. Salvucci & Taatgen (2008) Kieras & Meyer (1997) Kotseruba & Tsotsos (2020)
Research on Artificial Sapiens and artificial psychological organization can draw on these methods by defining stable observable processes and longitudinal conditions for comparison. The philosophical classification of an artificial agent and the psychological evidence for a particular cognitive mechanism are different questions; both can be investigated through explicit concepts and documented behavior. An operational program would test memory continuity, adaptive control, transfer, contextual consistency, and coordination with humans across changing tasks, while recording system updates and environmental supports. The useful scientific outcome is a reproducible account of functional organization and its constraints, with theoretical interpretation proportionate to the tested phenomena. Newell (1990) Ritter et al. (2019) Kotseruba & Tsotsos (2020)
Critical Scientific Synthesis and Open Problems
The reviewed architectures share a commitment to explaining multiple cognitive activities through reusable computational mechanisms. ACT-R makes commitments about modular coordination and retrieval; Soar emphasizes structured decision cycles and learned knowledge; EPIC stresses temporal concurrency and scheduling; CLARION integrates explicit and implicit representation; neural architectures model additional implementation constraints. These differences are scientifically productive because they suggest distinct experimental interventions. No single successful simulation validates every architectural claim, and evidence supporting one component should be distinguished from evidence for the whole framework. The most persuasive results are those that jointly explain behavior, learning, timing, and independent constraints without revising the common mechanism for each task. Anderson et al. (2004) Laird et al. (1987) Meyer & Kieras (1997) Sun & Zhang (2006) Eliasmith et al. (2012)
Several open problems organize the next phase of research. Architectural identifiability remains limited by flexible task programming and overlapping predictions. Neural validation requires more discriminating mappings between model operations and measurements. Developmental modeling requires realistic histories of experience, and individual-difference studies need repeated observations capable of separating parameters from strategies. Ecological tasks introduce interruptions, social coordination, and external supports that can alter apparent limits. Contemporary learned systems broaden the implementation space while making training provenance and interpretability increasingly important. These are linked problems of scientific design because each asks what must remain stable for an explanation to count as unified. Poldrack (2006) Taatgen & Anderson (2002) Kotseruba & Tsotsos (2020)
A cumulative research strategy should register architectural predictions and versioned code, provide stimulus and task specifications, use held-out conditions and new populations, and report model-recovery diagnostics. It should compare interpretable alternatives and retain a permanent record of failed contrasts. These procedures support a distinction between an architecture that can imitate observed cognition and one that explains why a defined family of human processes behaves as it does under intervention. The reviewed primary investigations provide important local evidence and demonstrations, but the general problem of transportable cognitive integration remains open to further empirical resolution. This conclusion follows from the diversity of research commitments and tests reviewed, rather than from an assumption that theoretical plurality itself is a defect. Newell (1990) Kotseruba & Tsotsos (2020) Ritter et al. (2019)
Conclusion
Cognitive architectures investigate how perception, memory, goals, learning, and action can be organized through a stable family of computational processes. The history from symbolic information-processing theories through ACT-R, Soar, EPIC, CLARION, and large-scale neural simulations reveals a sustained scientific effort to replace isolated task descriptions with integrated explanations. The explanatory value of a proposed architecture lies in its operational restrictions: what it permits, what it predicts, and which observations would contradict it. Foundational books and experiments establish the main proposals, while comparative reviews disclose the breadth and heterogeneity of the field. Newell (1990) Anderson et al. (2004) Laird (2012) Kotseruba & Tsotsos (2020)
The strongest empirical contributions include precise predictions for retrieval and action timing, experimental challenges to rigid interpretations of multitasking bottlenecks, models of structured task-set learning, and attempts to connect cognitive processes with neural activity. These results support particular mechanisms in particular conditions. Architectural universality requires an additional evidentiary step: genuinely independent cross-task predictions using common parameters, transparent strategy assumptions, reproducible software, and discriminating tests against competitors. This standard gives equal attention to successful explanation, error prediction, and informative failure. Meyer & Kieras (1997) Schumacher et al. (2001) Collins & Frank (2013) Eliasmith et al. (2012)
The unifying methodological result is that a cognitive architecture becomes an empirically meaningful theory when its enduring organization constrains observations across changing tasks and circumstances. This criterion is applicable to human cognition, neural simulation, artificial agents, and human–artificial interaction. It preserves the scientific distinction between demonstrated functional capability and explanations of human psychological organization while enabling rigorous comparisons among them. Future work will advance the field by making architectures harder to adjust after the fact and easier to test under independent experimental conditions. Such an approach realizes the original aspiration of unified cognition as a cumulative and revisable research program. Newell (1990) Ritter et al. (2019) Kotseruba & Tsotsos (2020)
References
- Anderson, J. R. (1983). The architecture of cognition. Harvard University Press. https://books.google.com/books/about/The_Architecture_of_Cognition.html?id=1OtfQgAACAAJ
- Anderson, J. R., Bothell, D., Byrne, M. D., Douglass, S., Lebiere, C., & Qin, Y. (2004). An integrated theory of the mind. Psychological Review, 111(4), 1036–1060. https://doi.org/10.1037/0033-295X.111.4.1036
- Baddeley, A. (2000). The episodic buffer: A new component of working memory? Trends in Cognitive Sciences, 4(11), 417–423. https://doi.org/10.1016/S1364-6613(00)01538-2
- Badre, D., & Frank, M. J. (2012). Mechanisms of hierarchical reinforcement learning in cortico-striatal circuits 2: Evidence from fMRI. Cerebral Cortex, 22(3), 527–536. https://doi.org/10.1093/cercor/bhr117
- Botvinick, M. M., Braver, T. S., Barch, D. M., Carter, C. S., & Cohen, J. D. (2001). Conflict monitoring and cognitive control. Psychological Review, 108(3), 624–652. https://doi.org/10.1037/0033-295X.108.3.624
- Collins, A. G. E., & Frank, M. J. (2013). Cognitive control over learning: Creating, clustering, and generalizing task-set structure. Psychological Review, 120(1), 190–229. https://doi.org/10.1037/a0030852
- Cooper, R., & Shallice, T. (2000). Contention scheduling and the control of routine activities. Cognitive Neuropsychology, 17(4), 297–338. https://doi.org/10.1080/026432900380427
- Daily, L. Z., Lovett, M. C., & Reder, L. M. (2001). Modeling individual differences in working memory performance: A source activation account. Cognitive Science, 25(3), 315–353. https://doi.org/10.1016/S0364-0213(01)00039-8
- Eliasmith, C., Stewart, T. C., Choo, X., Bekolay, T., DeWolf, T., Tang, Y., & Rasmussen, D. (2012). A large-scale model of the functioning brain. Science, 338(6111), 1202–1205. https://doi.org/10.1126/science.1225266
- Frank, M. J., & Badre, D. (2012). Mechanisms of hierarchical reinforcement learning in corticostriatal circuits 1: Computational analysis. Cerebral Cortex, 22(3), 509–526. https://doi.org/10.1093/cercor/bhr114
- Kieras, D. E., & Meyer, D. E. (1997). An overview of the EPIC architecture for cognition and performance with application to human-computer interaction. Human–Computer Interaction, 12(4), 391–438. https://doi.org/10.1207/s15327051hci1204_4
- Kotseruba, I., & Tsotsos, J. K. (2020). 40 years of cognitive architectures: Core cognitive abilities and practical applications. Artificial Intelligence Review, 53, 17–94. https://doi.org/10.1007/s10462-018-9646-y
- Laird, J. E., Newell, A., & Rosenbloom, P. S. (1987). SOAR: An architecture for general intelligence. Artificial Intelligence, 33(1), 1–64. https://doi.org/10.1016/0004-3702(87)90050-6
- Laird, J. E. (2012). The Soar cognitive architecture. MIT Press. https://doi.org/10.7551/mitpress/7688.001.0001
- Marr, D. (2010). Vision: A computational investigation into the human representation and processing of visual information (reissue of 1982 original). MIT Press. https://doi.org/10.7551/mitpress/9780262514620.001.0001
- Meyer, D. E., & Kieras, D. E. (1997). A computational theory of executive cognitive processes and multiple-task performance: Part 1. Basic mechanisms. Psychological Review, 104(1), 3–65. https://doi.org/10.1037/0033-295X.104.1.3
- Miller, G. A., Galanter, E., & Pribram, K. H. (1960). Plans and the structure of behavior. Holt, Rinehart and Winston. https://wellcomecollection.org/works/rbux7wwu
- Newell, A. (1990). Unified theories of cognition. Harvard University Press. https://search.worldcat.org/title/Unified-theories-of-cognition/oclc/21228315
- Poldrack, R. A. (2006). Can cognitive processes be inferred from neuroimaging data? Trends in Cognitive Sciences, 10(2), 59–63. https://doi.org/10.1016/j.tics.2005.12.004
- Ritter, F. E., Tehranchi, F., & Oury, J. D. (2019). ACT-R: A cognitive architecture for modeling cognition. WIREs Cognitive Science, 10(3), e1488. https://doi.org/10.1002/wcs.1488
- Rosenbloom, P. S., Laird, J. E., Newell, A., & McCarl, R. (1991). A preliminary analysis of the Soar architecture as a basis for general intelligence. Artificial Intelligence, 47(1–3), 289–325. https://doi.org/10.1016/0004-3702(91)90057-Q
- Salvucci, D. D., & Taatgen, N. A. (2008). Threaded cognition: An integrated theory of concurrent multitasking. Psychological Review, 115(1), 101–130. https://doi.org/10.1037/0033-295X.115.1.101
- Schumacher, E. H., Seymour, T. L., Glass, J. M., Fencsik, D. E., Lauber, E. J., Kieras, D. E., & Meyer, D. E. (2001). Virtually perfect time sharing in dual-task performance: Uncorking the central cognitive bottleneck. Psychological Science, 12(2), 101–108. https://doi.org/10.1111/1467-9280.00318
- Sun, R., & Zhang, X. (2006). Accounting for a variety of reasoning data within a cognitive architecture. Journal of Experimental & Theoretical Artificial Intelligence, 18(2), 169–191. https://doi.org/10.1080/09528130600557713
- Taatgen, N. A., & Anderson, J. R. (2002). Why do children learn to say ‘broke’? A model of learning the past tense without feedback. Cognition, 86(2), 123–155. https://doi.org/10.1016/S0010-0277(02)00176-2
- Tenenbaum, J. B., Kemp, C., Griffiths, T. L., & Goodman, N. D. (2011). How to grow a mind: Statistics, structure, and abstraction. Science, 331(6022), 1279–1285. https://doi.org/10.1126/science.1192788