Preview

Baseball Quiz

• the point is that the correct answer depends on knowledge about baseball

Function of a Cognitive Architecture

Cognitive agent: a function to mapping history to actions. Precepts -> Action

Levels of Cognitive Architectures

1. Hardware /Implementation level
2. Algorithm/Symbol Level
3. Task / knowledge Level

• levels are interconnected. the lower level provides architecture for higher level and higher level provides content to the lower level. The constrain each other and some time determine each other.
• Most of the AI work deals with Level 2 and level 3.

Quiz three layers of Watson

a physical computer, algorithm contains at least searching function. Task is usually answering certain questions.

Assumptions of Cognitive Architectures

Architecture + Content = Behavior

• Architecture + Content = Behavior
• now, assume we can fix the architecture, then behaviour change then indicates Content change.
• Now given that function of cognitive architecture, then we can say if action changes, then the knowledge change must be the cause. This is a simplification of the understanding of human cognitive.

A Cognitive Architecture for Production Systems

SOAR

• Episodic: event
• Semantic: generalizations, concept
• Procedural: how to do certain things

Return to the Pitcher

How AI make decisions on the pitcher problem

• the pitcher has external knowledge: the status of the game of, other players' position...
• the pitcher also has internal knowledge: his goal

Action Selection

Actions Abstract state planning

Putting Content in the Architecture

Paste_Image.png

Bringing in Memory

Paste_Image.png Production Rules as Procedure Memory

• the internal and external knowledge are in working memory
• procedural rules are long-term memory

Exercise: Production System in Action I

• we mesh the knowledge in working memory (percept) and rules in the procedural type of long-term memory and make the decision on what action to take. This is a P -> A mapping

Exercise: Production System in Action II

• r4 suggests that working memory and long-term memory interact with each other constantly, and new rules,new knowledge and actions can be generated in the process.

• If a new left-handed batter, then two rules can be possible.
• the problem of no action is that there no rule for this particular state.

Chunking

When 2 operators were selected, there is now rule to determine what to do. Then SOAR will have to learn a rule by activating the episodic memory.

Chunkin is a learning procedure. (for details, see Lehman, J. F., Laird, J. E., & Rosenbloom, P. S. (1996). A gentle introduction to Soar, an architecture for human cognition. Invitation to Cognitive Science, 4, 212-249.)

chunking

Chunking in operation.

Fundamentals of Learning

How do agent learn? when to learn, what to learn, why to learn.

Assignment

Recap

Paste_Image.png

The Cognitive Connection

• working memory in both human and production systems have capacity limitations
• Soar and human have similarity when solving algebra problems

20170202 first draft