Cognitive Architectures

Цели урока

• Understand SOAR principles: problem space, impasse, chunking as online learning
• Master ACT-R math: activation, retrieval time, power law of forgetting
• Understand LIDA: conscious broadcast as an attention mechanism
• Compare architectures and recognize their ideas in modern LLM agents

Предварительные знания

• Global Workspace Theory (lesson 11)
• Basic understanding of IF-THEN production systems
• Intuitive familiarity with neural networks

SOAR flew a fighter jet in 1995. ACT-R predicts fMRI patterns. LIDA engineers GWT. All before LLMs. Modern agents reinvent these ideas from scratch.

• TacAir-Soar (1995): SOAR controlled a virtual F-16 in US Air Force simulators
• ACT-R (2004): model reproduced fMRI patterns during math tasks down to specific brain regions
• LIDA is used to model attention and disorders of consciousness in neuroscience
• ReAct, AutoGPT, LangGraph: implicitly reinvent SOAR (planning) and ACT-R (memory)
• Transformer attention is a partial implementation of GWT, but without explicit winner-take-all

Three Paths to General Intelligence

Allen Newell and John Laird created SOAR in 1983 as an attempt to realize Newell's 'unified theory of cognition'. John Anderson developed ACT-R in 1993, emphasizing mathematical precision and behavior prediction. Stan Franklin created LIDA in 2005, integrating GWT. All three worked in parallel - different views on the same problem.

SOAR: Learning Through Impasse

**1995. TacAir-Soar controls a virtual F-16 in a US Air Force simulator. The system flies, evades missiles, engages targets - and learns while doing it.** Not a neural network, not an LLM. A production system with 50,000 IF-THEN rules that creates new rules through chunking every time it hits an impasse. This is SOAR - State, Operator And Result.

**SOAR** is a problem-oriented architecture by Allen Newell and John Laird (1983). All cognition is movement through problem space: current state + operator = new state. When no operator can be selected, an impasse occurs, a subgoal is created, and after resolution - a new rule (chunk) is generated. Chunking = compilation of experience.

Three Memory Types and the Main Cycle

ACT-R: Mathematics of Human Memory

**ACT-R (Adaptive Control of Thought - Rational) does something no neural network can: predict human reaction time to the millisecond.** In 2004, an ACT-R model accurately reproduced fMRI patterns during mathematical problem solving - down to specific brain regions. John Anderson created not just an AI architecture, but a formal model of human cognition.

**ACT-R central bottleneck**: only ONE production rule can execute per cycle (~50 ms). This explains the psychological refractory period - why humans cannot fully parallelize two tasks. Peripheral modules (vision, hearing, hands) work in parallel, but the central processor is strictly sequential.

The Memory Activation Formula

The key idea in ACT-R: each chunk (unit of knowledge) has **activation A = B + S + e**, where B is base-level (usage frequency), S is spreading activation from context, e is noise. Retrieval time: **T = F * exp(-A)**. Higher activation means faster retrieval.

LIDA: Consciousness Through Coalition Competition

**LIDA (Learning Intelligent Distribution Agent) is the only cognitive architecture that explicitly implements Global Workspace Theory.** Stan Franklin created it in 2005 as an engineering realization of Baars' theory: consciousness is the winner of competition for the global workspace. Each cognitive cycle (~200-500 ms) is a competition among coalitions for the right to be broadcast to all system modules.

**Codelets** are small specialized processes inside LIDA. Attention codelets search for what is important, behavior codelets propose actions, structure building codelets construct understanding from sensory data. This implements Marvin Minsky's 'society of mind' idea - intelligence as interaction of simple agents.

Comparison and Application to LLM Agents

**All three architectures solve the same problem in different ways: how to organize intelligence that not only reacts but plans, learns, and remembers.** Modern LLM agents reinvent the same components - often worse, because they lack 40 years of cognitive science behind them.

How These Ideas Live in LLM Agents

When designing an LLM agent, it pays to implement explicitly: Working Memory (what is in focus now), Goal Stack (goal hierarchy), Episodic Memory (what happened), Meta-Cognition (confidence monitoring). Systems that add these components explicitly are more stable than those where everything is implicit in parameters.

Вопросы для размышления

• When designing an LLM agent for a real task - which components from SOAR/ACT-R/LIDA are worth implementing explicitly, and why cannot one rely on them being emergent from parameters?

Связанные уроки

• arch-04-cpu