# Cognitive Architecture A cognitive architecture is a theory of the fixed structures and processes that underlie cognition—the invariant framework within which learning, reasoning, and problem-solving occur. While knowledge varies across individuals and tasks, the architecture remains constant, defining what kinds of representations the mind can have, how they're processed, and what computational resources are available. Major cognitive architectures like ACT-R ([[John R. Anderson]]), SOAR ([[Allen Newell]]), and Global Workspace Theory ([[Bernard Baars]]) attempt to unify findings across perception, memory, attention, and reasoning into coherent computational systems. Cognitive architectures operationalize the [[Computational Theory of Mind]], specifying not just that the mind computes, but *how* it computes. They draw on [[Jerry Fodor]]'s [[Modularity of Mind]] (specialized input systems) while attempting to model central cognition (general reasoning). These architectures serve both as scientific theories (tested against human data) and engineering frameworks (implemented in AI systems). The challenge remains integrating insights from symbolic approaches with neural/connectionist models—leading to hybrid architectures that combine rule-based reasoning with learned representations. ## Architecture Overview ``` ┌─────────────────────────────────────────────────────────────┐ │ GENERIC COGNITIVE ARCHITECTURE │ ├─────────────────────────────────────────────────────────────┤ │ │ │ PERCEPTION ACTION │ │ ┌─────────┐ ┌─────────┐ │ │ │ Vision │ │ Motor │ │ │ │ Hearing │ │ Speech │ │ │ │ etc. │ │ etc. │ │ │ └────┬────┘ └────▲────┘ │ │ │ │ │ │ ▼ │ │ │ ┌─────────────────────────────────────────────────────┐ │ │ │ WORKING MEMORY │ │ │ │ (limited capacity, temporary) │ │ │ └───────────────────────┬─────────────────────────────┘ │ │ │ │ │ ┌────────────┴────────────┐ │ │ ▼ ▼ │ │ ┌─────────────────┐ ┌─────────────────┐ │ │ │ DECLARATIVE │ │ PROCEDURAL │ │ │ │ MEMORY │◄─────▶│ MEMORY │ │ │ │ (facts, events)│ │ (skills, rules) │ │ │ └─────────────────┘ └─────────────────┘ │ │ │ │ ┌─────────────────────────────────────────────────────┐ │ │ │ CENTRAL PROCESSOR │ │ │ │ • Production rules / inference │ │ │ │ • Goal management │ │ │ │ • Conflict resolution │ │ │ └─────────────────────────────────────────────────────┘ │ │ │ └─────────────────────────────────────────────────────────────┘ ``` ## Major Architectures | Architecture | Developer | Key Features | |--------------|-----------|--------------| | **ACT-R** | [[John R. Anderson]] | Declarative/procedural memory, production rules | | **SOAR** | [[Allen Newell]] | Problem spaces, chunking, universal subgoaling | | **EPIC** | Meyer & Kieras | Perceptual-motor focus, parallel processors | | **CLARION** | Ron Sun | Implicit/explicit dual representation | | **Global Workspace** | [[Bernard Baars]] | Conscious broadcast, competition for access | | **LIDA** | Stan Franklin | GWT + action selection | ## ACT-R Components | Component | Function | |-----------|----------| | **Declarative module** | Stores facts (chunks) | | **Procedural module** | Stores skills (production rules) | | **Goal module** | Tracks current intentions | | **Imaginal module** | Problem representation | | **Visual module** | Visual processing | | **Motor module** | Action execution | | **Production system** | IF-THEN rule matching | ## SOAR Principles | Principle | Description | |-----------|-------------| | **Problem spaces** | All tasks as search through states | | **Universal subgoaling** | Impasses trigger new goals | | **Chunking** | Learn new rules from experience | | **Working memory** | Holds current situation | | **Long-term memory** | Procedural, semantic, episodic | ## Symbolic vs Connectionist vs Hybrid | Aspect | Symbolic | Connectionist | Hybrid | |--------|----------|---------------|--------| | **Representation** | Explicit symbols | Distributed weights | Both | | **Learning** | Rule induction | Gradient descent | Multiple mechanisms | | **Reasoning** | Logical inference | Pattern completion | Integrated | | **Examples** | ACT-R, SOAR | PDP models | CLARION, ACT-RN | ## Cognitive Constraints Modeled | Constraint | Architectural Feature | |------------|----------------------| | **Limited attention** | Working memory capacity | | **Serial bottleneck** | Central processor limitations | | **Practice effects** | Strengthening mechanisms | | **Forgetting** | Decay, interference | | **Transfer** | Production generalization | ## Applications | Domain | Application | |--------|-------------| | **HCI** | Predicting user behavior | | **Education** | Intelligent tutoring systems | | **AI** | Cognitive agents | | **Psychology** | Testing theories computationally | | **Neuroscience** | Linking to brain regions | ## Relation to [[Modularity of Mind]] | Fodor's View | Cognitive Architectures | |--------------|------------------------| | Input systems modular | Perceptual modules accepted | | Central cognition non-modular | Attempt to model anyway | | Modules encapsulated | Some information sharing | | Domain-specific | Mix of general/specific | ## Open Questions | Question | Challenge | |----------|-----------| | **Unified vs modular** | How integrated is cognition? | | **Symbolic vs subsymbolic** | What level of abstraction? | | **Learning** | How are new skills acquired? | | **Consciousness** | How does awareness arise? | | **Embodiment** | Role of body and environment? | ## References - Anderson, John R. *The Architecture of Cognition* (1983) - Newell, Allen. *Unified Theories of Cognition* (1990) - Baars, Bernard. *A Cognitive Theory of Consciousness* (1988) - https://en.wikipedia.org/wiki/Cognitive_architecture ## Related - [[Computational Theory of Mind]] - [[Modularity of Mind]] - [[Jerry Fodor]] - [[John R. Anderson]] - [[Allen Newell]] - [[Working Memory]] - [[Long-Term Memory]] - [[Global Workspace Theory]] - [[Cognitive Science]]