John Robert Anderson, working at Carnegie Mellon University, developed ACT-R (Adaptive Control of Thought-Rational), the most comprehensive and widely used cognitive architecture in psychology. ACT-R integrates declarative memory, procedural memory, perceptual-motor modules, and goal management within a unified framework grounded in rational analysis -- the principle that cognition is optimally adapted to the statistical structure of the environment.
ACT-R Architecture
t_j = time since the j-th use of chunk i
d = decay parameter (approximately 0.5)
beta_i = base-level constant
Probability of retrieval: P = 1/(1 + exp(-(B_i - tau)/s))
ACT-R distinguishes between declarative memory (factual knowledge stored as chunks) and procedural memory (skills stored as production rules). The activation of a declarative chunk determines its accessibility: activation reflects both base-level strength (frequency and recency of use, following a power law of forgetting) and spreading activation from associated chunks in the current context. When activation exceeds a threshold, the chunk can be retrieved; retrieval time decreases with increasing activation.
Anderson's rational analysis framework (1990) proposes that cognitive mechanisms are optimally adapted to the statistical structure of the environment. Memory decay follows a power function because the probability of needing information decays as a power function of time. Categorization uses exemplar-like representations because the environment contains cluster structure. This "function-first" approach aligns with Marr's computational level of analysis.
Production System and Learning
Procedural knowledge in ACT-R is represented as condition-action rules (productions). When a production's conditions match the current state of buffers, it fires and executes its action. Production learning occurs through compilation -- the merging of sequential productions into single, more efficient rules. Utility learning uses a reinforcement-like mechanism where productions that lead to goals receive increased utility. This dual learning system accounts for the transition from slow, declarative problem solving to fast, automatized skill execution.
Legacy and Impact
ACT-R has been applied to hundreds of tasks including arithmetic, algebra, programming, language comprehension, driving, and air traffic control. Its precise quantitative predictions about response times, error patterns, and brain activation (through module-to-brain-region mappings) have made it a uniquely productive framework for both basic research and practical applications in intelligent tutoring systems and human-computer interaction.