280 likes | 306 Views
Delve into cognitive architectures, multi-agent systems, and blackboard systems in AI. Discover ICARUS architecture and memory representations in building intelligent systems.
E N D
Cognitive Architectures and General Intelligent Systems Pay Langley 2006 Presentation : Suwang Jang
Index • A trend of AI • Original Goal of AI and modern AI • Three Architectural Paradigms • Multi-agent systems • Blackboard systems • Cognitive Architecture • Commitments of Cognitive Architecture • ICARUS Architecture • Memories and Representations • Performance and Learning Process
Envision by early AI researchers • The Original Goal of AI was constructing artifacts which have almost same intellectual capacity as humans ☞ General Intelligent Systems
Computer vision • Computational linguistics • Planning • ……. But, Modern AI? - Fragmented Approaches -
? !! !! ? !! Newell’s arguments (1973) • He was critiquing the strategy of experimental cognitive psychologists, who studied isolated components of human cognition without considering their interaction • And he argued that we should evaluate AI in terms of generality and flexibility, rather than success on a single domain
The Notion of Newell • Cognitive Psychology + (as close allies) • AI Research ☞ “Cognitive Architecture” (1973)
Three Architectural Paradigms for General Intelligent System • Multi-agent System • Blackboard System • ACT, Soar and ICARUS (Cognitive Architecture Based)
Multi-agent System (Sycara 1998) • Traditional approaches to software engineering • Features • Distinct modules • Direct communication with each other (Specified Input/Output and Protocol) • No constraints on how each module operates • Advantage • Easy for teamwork (Developing each module separately and Integrating them) • Disadvantage • Need for modules to communicate directly with one another
Pattern matching against elements Blackboard System(Engelmore and Morgan 1989) • Retains Modularity of the first framework • Indirect Communication through short-term memory • More closer to theories of human cognition
Newell’s View • Unified theory of intelligent behavior, not simply integrated one • Mutual constraints for independency among modules • Architectural design changed only gradually for correspondence to new structure that supporting new functionality
Cognitive Architecture • The short-term and long-term memories that store the agent’s beliefs, goals, and knowledge • The representation and organization of structures that are embedded in these memories • The functional processes that operate on these structures, including both performance and learning mechanisms • A programming language that lets one construct knowledge-based systems that embody the architecture’s assumption
The ICARUS Architecture • Common cognitive architecture + concern with physical agent that operate in an external environment
Principles • Cognition is grounded in perception and action • Concepts and skills are distinct cognitive structures • Long-term memory is organized in a hierarchical fashion • Skill and concept hierarchies are acquired in a cumulative manner • Long-term and short-term structures have a strong correspondence
Memories and Representations ④ ① ③ ② ⑤
① Conceptual Memory • Concept : • Head ☞ name arguments • Body • :percept ☞ type, attribute value (from Perceptual Buffer) • :relation ☞ low-level concept • :test (primitive concept) ☞ Boolean test • Bottom-up
Non-Primitive Non-Primitive Primitive Primitive Long-term concept memory
② Skill Memory • Primitive skill : • Head ☞ Concept which the clause should achieve upon successful completion • Body • :start ☞ describe the situation in which the agent initiate the clause • :require ☞ field that must hold throughout execution • :actions ☞ executable action (to Motor Buffer)
② Skill Memory • Non-primitive skill : • No :require field and :action field • Instead have a :subgoals field • Top-down
Non-Primitive Non-Primitive Non-Primitive Recursive Call Primitive Non-Primitive Primitive Non-Primitive Long-term skill memory
Short-term Memory • ③ Belief memory • (Concept name + Instance) • ④ Perceptual buffer • (type, unique name, attribute + value …) • ⑤ Goal/Intention Memory • Stack of goals … Sub goal Sub goal High-level goal
Skill Clause • Top-down manner • If execution module can find an applicable path, it carry out actions. • Applicable path : • Concept instance of goal is not satisfied yet • Requirements of terminal skill are satisfied • For each skill instance in the path not executed on the previous cycle • The start conditions are satisfied
Skill Clause • If execution module can not find applicable path It evokes a module for Means-ends problem solving (Newell and Simon 1961) • Push new goals and concept definition needed to achieve top-level goal onto goal stack until it find one it can achieve with an applicable skill • Applicable skill -> pop • Unsatisfied concept -> push sub-concepts • If none remain -> pop the parent • This processes continues until system achieve top-level goal
Learning • A learning module creates a new skill whenever problem solving • Achieved goal + subgoals as subskills + start condition • It discussed in more detail elsewhere (Langley and Choi, 2006)
Simulation of In-city driving • ICARUS program for delivering packages in simulated driving environment • Simulated environment • buildings, road segments, intersections, lane lines, packages, other vehicles, and agents’ vehicles • 15 primitive concepts and 55 higher-level concepts (6 level deep) • 8 primitive skills and 33 higher-level skills (5 level deep) • Result : Changing speed, altering wheel angle, depositing packages