Cognitive architectures and rational analysis: the search for one mind

Newell (1990) argued that cognitive science was ready for unified theories, and offered one; Anderson (1983) built another. The promise was that a fixed set of mechanisms would reproduce skilled behaviour across domains. This second of ten units argues that the durable contribution is the modelling machinery, not the unification it promised, and it is the last to keep the explanation inside one head. An architecture of the individual mind has no terms for what is spread across person, artefact and setting. Its applications arrived with the task and the interface supplied from outside.

The promise: one mind across every band of behaviour

Newell (1990) set the ambition and its test. The case for unified theories rested on scope rather than precision. A candidate had to account for the full range of human cognition, as one system rather than a collection of models. Anderson (2002) states the demand in the terms Newell proposed, the biological, cognitive, rational and social bands. He measures the gap: cognitive psychology studies effects in tens of milliseconds, while significant educational outcomes take tens of hours.

Anderson (2002) assesses the bet as three claims. The Decomposition Thesis holds that learning at the social band reduces to learning at lower bands. The Relevance Thesis holds that instructional outcomes improve when lower-band cognition is respected. The Modelling Thesis holds that cognitive modelling bridges the scales. His verdict is uneven: decomposition has good evidence down to about ten milliseconds, the other two less.

What the architecture contains, and what it leaves outside

An architecture is a claim about what is invariant. Anderson (1983) set out ACT* as the basic principles of operation built into the cognitive system. They run from production systems to memory for facts and procedural learning. Laird, Newell and Rosenbloom (1987) claimed the same for a rival design. Their SOAR is an implemented architecture meant to support the full range of cognitive tasks, problem solving methods and representations, and to learn about every aspect of a task.

The consequence for interaction research is an argument rather than a finding, because no architecture paper states it. Both describe one agent and treat the task as something that agent is given. Neither has a term for how work is divided between a person, a display and a second person.

Rational analysis: the environment does the fitting

Anderson (1990) supplied the programme's most distinctive move. Rational analysis predicts behaviour from the structure of the environment rather than from a model of internal mechanisms. Specify the system's information-processing goals, the environment's structure and the computational constraints, work out what behaviour would be optimal, and the prediction follows. He applied it to memory, categorisation, causal inference and problem solving.

The empirical heart of the method is Anderson and Schooler (1991). They examined three environmental sources: the New York Times, parental speech and electronic mail. In 730 days of headlines, from 1 January 1986 to 31 December 1987, the probability that a word was needed on a later day rose with the number of times it had appeared in the previous 100 days, and the same relationship held in the other two sources. Memory performance tracked the same statistics of frequency, recency and prior exposure. They argued that memory has the form it does because it is adapted to those environmental relationships.

The move buys the architecture predictive power over effects that need not be fitted one at a time. The architecture's parameters become a claim about the statistics of one setting, that of readers at home in English. A setting is what an account of an individual mind does not supply.

The mature architecture integrates its modules around a single rule

Anderson and colleagues (2004) state the mature theory. ACT-R's perceptual-motor, goal and declarative memory modules are specialised systems in distinct cortical regions. They place chunks in buffers, and a production system responds to the patterns in those buffers. At any moment the system selects a single production rule. The authors offer that serial bottleneck as the source of organised control, and the empirical examples require little parameter estimation.

EPIC was built for interaction rather than for psychology. Kieras and Meyer (1997) describe peripheral sensory-motor processors surrounding a production-rule cognitive processor, designed for multimodal and multiple-task performance. They report models giving precise accounts of performance speed at the interface. Multiple-task performance there rests on adaptive executive control, so the programme could reach concurrency when an architecture was designed for it.

Each demonstration had to be given what the architecture lacks

Byrne (2001) applied ACT-R/PM, the architecture extended with perceptual and motor modules, to Nilsen's random menu selection task. His position is that understanding a graphical interface requires the user's cognitive, perceptual and motor capabilities, together with the task and the artefact. Guided by an eye-tracking study of 11 participants, which found mean response time tracking mean fixations at a correlation of 0.97, he built a series of models. The task was completely specified: find and click a digit in a randomly ordered menu, at roughly 100 milliseconds for each further position down. The final model gives a zero-parameter fit he calls excellent though not perfect, and the fixations showed local decisions rather than a strategy fixed in advance.

Howes and Young (1996) offer the learning case. Their model, TAL, is built in Soar with a single set of architectural mechanisms and learns task-action mappings. It reproduces the human preference for mappings that are consistent, interactive and meaningful. The authors' own limits: TAL is no powerful learner, and it does not model error recovery.

Both demonstrations had to be given something the architecture does not contain. Byrne's model took its perceptual guidance from that study. The consistency of the device's mapping sets TAL's difficulty, and that property of the interface is described rather than derived. The architecture predicts what an agent does with an arrangement, and the arrangement arrives from outside it.

Contested question: has the programme unified, or only multiplied?

One position holds that unification was achieved. Anderson and colleagues (2004) present a theory whose modules are integrated rather than merely coordinated. Its predictions cover choice, latency and brain data at little cost in parameter fitting. Laird, Lebiere and Rosenbloom (2017) report a different convergence: a standard model of the mind. It began as a consensus at the 2013 AAAI Fall Symposium and was extended by synthesis across ACT-R, Sigma and Soar. On this account the promise has largely been kept.

The other position holds that multiplication is the finding. Kotseruba and Tsotsos (2020) surveyed the field's first forty years and found 84 architectures in their final set, 49 still actively developed, against an estimated three hundred of which at least a third are active. Lieto, Lebiere and Oltramari (2018) argue that the representational level of several of the most used architectures is under-specified in the size and the type of the knowledge encoded, which limits how far a comparison with human knowledge can be pushed. Anderson (2002) adds a doubt from inside: the modelling bridge across the bands awaits more work. On this account the standard model records agreement about components, and is not a theory that generates the predictions the programme promised.

Counting architectures will not settle it. What would settle it is a comparison neither side has run. Its design is benchmark tasks across several bands and domains, with each architecture's predictions made before the data are seen and without refitting, which separates the transfer of parameters from fitting them afresh. A second test would settle the meta-question: how much agreement on modules counts as a theory. Until then the convergence claim cannot fail, and multiplication cannot be read as failure either.

The evidence differs in kind on each side. The convergence claim rests on fit within one architecture and on a consensus document agreed by the architects themselves, which is agreement rather than a test. The multiplication claim rests on a census of 84 architectures, which counts what exists without measuring what any of them explains. While the dispute is open, a designer should treat an architecture claim as a hypothesis about one user on one specified task, and should request the benchmark before any transfer is trusted. The strongest objection is to the unit's claim that the modelling machinery is the durable contribution. Lieto and colleagues (2018) argue that the knowledge level of the main architectures is under-specified. If so, what travels between projects is a vocabulary of modules and buffers rather than a working instrument.

Where the argument is heading

Three lines are live. The first continues the standard-model project: Laird, Lebiere and Rosenbloom (2017) present it as a cumulative reference point that names its own remaining incompleteness. The second repairs the representational level that Lieto, Lebiere and Oltramari (2018) identify, through semantic pointers, conceptual spaces and hybrid neuro-symbolic approaches, which Lieto and colleagues judge still unproven. The third takes the vocabulary where the programme never went. Sumers, Yao, Narasimhan and Griffiths (2024) set out CoALA, a language agent with modular memory components, a structured action space and a general decision procedure, built on the history of cognitive architectures. CoALA is a taxonomy rather than an implemented architecture: it claims no operator times, makes no prediction before a task and is not compared against human data, so the programme's central bet is absent from its most visible descendant.

The sources this unit rests on

• Anderson, J. R. (1983). The Architecture of Cognition. Harvard University Press. doi:10.4324/9781315799438 Sets out ACT* as the basic principles of operation built into the cognitive system. It is the origin of the programme's ambition.
• Laird, J. E., Newell, A. and Rosenbloom, P. S. (1987). SOAR: an architecture for general intelligence. Artificial Intelligence 33(1). doi:10.1016/0004-3702(87)90050-6 Presents SOAR as an implemented architecture for general intelligence, on the goal of a structure letting a system perform the full range of cognitive tasks and problem solving methods, and learn about every aspect of a task.
• Anderson, J. R. (1990). The Adaptive Character of Thought. Lawrence Erlbaum. doi:10.4324/9780203771730 Develops rational analysis, which predicts behaviour from the structure of the environment by specifying the system's goals and the environment's structure. It applies the method to memory, categorisation, causal inference and problem solving.
• Newell, A. (1990). Unified Theories of Cognition. Harvard University Press. https://archive.org/details/unifiedtheorieso0000newe Makes the case for unified theories of cognition and sets out Soar as a candidate, on the argument that such theories are within reach and worth attempting. It supplies the bands of cognition that later work uses to measure that reach.
• Anderson, J. R. and Schooler, L. J. (1991). Reflections of the environment in memory. Psychological Science 2(6). doi:10.1111/j.1467-9280.1991.tb00174.x Examined the New York Times, parental speech and electronic mail, and found that the need for a memory tracked frequency and recency as memory performance did. It argues that memory is adapted to those environmental relationships, the empirical heart of rational analysis.
• Howes, A. and Young, R. M. (1996). Learning consistent, interactive, and meaningful task-action mappings: a computational model. Cognitive Science 20(3). doi:10.1207/s15516709cog2003_1 Builds TAL in Soar with a single set of architectural mechanisms to learn task-action mappings, matching the human preference for mappings that are consistent, interactive and meaningful. The authors state that TAL is no powerful learner and does not model error recovery.
• Kieras, D. E. and 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). doi:10.1207/s15327051hci1204_4 Describes an architecture built for human-computer interaction, with peripheral processors around a production-rule cognitive processor, and reports models giving precise accounts of performance speed at the interface. Its multiple-task models rest on adaptive executive control.
• Byrne, M. D. (2001). ACT-R/PM and menu selection: applying a cognitive architecture to HCI. International Journal of Human-Computer Studies 55(1). doi:10.1006/ijhc.2001.0469 Applies ACT-R/PM to Nilsen's random menu selection task, on the position that interface use requires the user's capabilities, the task and the artefact together. Guided by eye-tracking data on visual search, the model ends in a zero-parameter fit Byrne calls excellent though not perfect.
• Anderson, J. R. (2002). Spanning seven orders of magnitude: a challenge for cognitive modelling. Cognitive Science 26(1). doi:10.1207/s15516709cog2601_3 Frames the scale problem as Newell's (1990) biological, cognitive, rational and social bands, and assesses decomposition, relevance and modelling as three theses. It finds good evidence for decomposition and unfinished work on the other two.
• Anderson, J. R., Bothell, D., Byrne, M. D., Douglass, S., Lebiere, C. and Qin, Y. (2004). An integrated theory of the mind. Psychological Review 111(4). doi:10.1037/0033-295x.111.4.1036 Presents the integrated ACT-R theory, whose modules place chunks in buffers for a production system that selects a single rule at any moment. Examples across choice, latency and brain data need little parameter estimation.
• Laird, J. E., Lebiere, C. and Rosenbloom, P. S. (2017). A standard model of the mind: toward a common computational framework across artificial intelligence, cognitive science, neuroscience, and robotics. AI Magazine 38(4). doi:10.1609/aimag.v38i4.2744 Proposes a standard model extended by synthesis across ACT-R, Sigma and Soar, spanning structure, memory, learning, and perception and motor. It names the loci of agreement and the remaining incompleteness, and it is the convergence side's best exhibit.
• Lieto, A., Lebiere, C. and Oltramari, A. (2018). The knowledge level in cognitive architectures: current limitations and possible developments. Cognitive Systems Research 48. doi:10.1016/j.cogsys.2017.05.001 Reviews the knowledge level of the most used architectures and finds two problems, the limited size and the homogeneous typology of the knowledge encoded, which limits comparison with human knowledge. It assesses semantic pointers, conceptual spaces and neuro-symbolic approaches as repairs.
• Kotseruba, I. and Tsotsos, J. K. (2020). 40 years of cognitive architectures: core cognitive abilities and practical applications. Artificial Intelligence Review 53(1). doi:10.1007/s10462-018-9646-y Surveys forty years of cognitive architectures and reports 84 in its final set, 49 still actively developed, against an estimated three hundred of which at least a third are active, with over 900 practical projects built on them. The multiplication is the unit's evidence against reading the programme as a convergence.
• Sumers, T. R., Yao, S., Narasimhan, K. and Griffiths, T. L. (2024). Cognitive architectures for language agents. Transactions on Machine Learning Research. doi:10.48550/arXiv.2309.02427 Sets out CoALA, a language agent with modular memory components, a structured action space and a general decision procedure, built on the history of cognitive architectures. It is a taxonomy, not an implemented architecture, so it carries the vocabulary without the predictive commitment.