Theoretical Foundations of Cognitive Architectures
Cognitive architectures provide frameworks for modeling human-like intelligence by integrating perception, memory, and reasoning into a unified system. These systems contrast with narrow AI by focusing on general-purpose problem solving.
Core Frameworks
- SOAR: A symbolic architecture that utilizes production rules and state-space search for complex problem-solving.
- ACT-R: A framework that combines symbolic production rules with sub-symbolic activation processes, modeling the latencies and accuracies of human cognition.
- OpenCog (GitHub): An open-source project developing a decentralized hypergraph-based model for general intelligence.
Legacy and Specialized Architectures
- General Problem Solver (GPS): A 1959 program by Newell and Simon, one of the first attempts to separate domain-specific knowledge from general-purpose reasoning strategies.
- CLARION: A cognitive architecture that emphasizes the distinction between implicit and explicit mental processes.
- LIDA (Learning Intelligent Distribution Agent): A comprehensive model of the mind based on Global Workspace Theory.
- Copycat / Metacat: Douglas Hofstadter's models of analogy-making and fluid concepts, focusing on the "slipnet" architecture for semantic flexibility.
Hybridization with Generative AI
Current research explores using LLMs as high-capacity associative memory modules within established cognitive architectures. This approach seeks to utilize LLMs for linguistic tasks while SOAR or ACT-R modules maintain long-term planning, logical consistency, and goal state monitoring.