Cognative Architecture

  • A cognative agent is a function f that maps a perceptual history P* to an action A.
  • Percepts -> Actions
  • The * stands for “history”
\[f: P^* \rightarrow A\]

Assumptions of Cognative Architectures:

  • Goal-oriented
  • Rich, complex environments
  • Significant knowledge
  • Represented at a level of abstraction, symbols
  • Respond flexibly to a changing environment
  • Learn from their experiences
\[\text{Architecture} + \text{Content} = \text{Behavior}\]

If architecture is fixed, then we only need to change the knowedge content of the agent to achieve different behaviors.

SOAR

SOAR

Consists of a long term memory portion

  • Procedural knowledge (how to pour water into a container)
  • Semantic knowledge (concepts and models of the environment, model of how a plane flies in the air)
  • Episodic knowledge (events, think what was for dinner yesterday)

and short term memory

  • Working memory

Production Systems

Levels of Abstraction:

  • High - Task / Knowledge Level
  • Mid - Algorithm / Symbol Level

  • Low - Hardware / Implementation Level

  • Low level provides architecture for higher levels up ladder
  • High level provides content for lower levels down ladder

Working Memory Cycle

  • Working memory is the state of the system
  • Long term memory can consist of “production rules”
  • Based on long-term memory some rules get activated
  • Rules which are activated consequences get established
  • Consequences that are established mutate working memory

If SOAR reaches a state that it cannot determine an action for, it will envoke episodic knowledge to try and learn a new behavior

Action Selection

Mapping percepts in the world into actions

  • Percepts from the world into a pitch selection

Chunking

  • A learning technique that SOAR uses to overcome impasses
  • Uses episodic memory to find an event in the past that is related to current percepts and leverages the previous event into a new rule

Misc

Reasoning first, then work backwards to learning