Symbols and abstractions flexible and function of the

This preview shows page 25 - 32 out of 52 pages.

Symbols and abstractionsFlexible and function of the environmentLearn from experienceArchitecture + Content = BehaviorIf the architecture is fixed, we only need to modify the content.We can map behavior to content because the architecture is fixed.Keep architecture constant, change content.So what is a good architecture?That’s next.A Cognitive Architecture for Production Systems
A specific architecture for deliberation: SOARSOAR consists of a long term memory and a working memory.The long term memory contains 3 types of knowledge:Episodic - events (what did you have for dinner yesterday)Semantic - generalizations in the form of concepts and models of the world (your model of howa plane flies)Procedural - how to do certain things (how to pour water from a jug into a tumbler)Return to the PitcherPitcher has to decide on an actionP* -> AHow will the pitcher as an AI agent come to this decision?How do the percepts (information) get mapped to the action to walk the batter?Pitcher has several kinds of knowledge - some internal (goals and objectives), some can beperceived from the environment (such as bases, state of game, batter, score, etc.)Action SelectionPitcher may look at various actions available and look at additional possibilities that appear aftermaking each decision. Sets up a state space.
Informal:Formal description:Can think of pitcher’s decision as an abstract state space and exploring the state space in an
attempt to map a path from the starting state to the goal state.Putting Content in the ArchitectureBringing in MemorySOAR’s working memory now contains all the properties described above that have beenmapped like so:
Some are percepts, others are pitcher’s internal goals.Imagine that the procedural part of SOAR’s long term memory contains rules.Sometimescalled “production rules” (the term “Production Systems” comes from this term). One of the firstthings the pitcher had to decide was whether to throw a pitch or walk a batter.There are rulesto account for this.SOAR’s long term memory consists of various types of knowledge.One type is procedural knowledge (about how to do something).
This knowledge is presented in the form of production rules of the form:If (something) then (something).There are antecedents and consequences, both can be connected with relationships like “and”and “or”.Another example where first rule is skipped due to runner on first:As the contents of the working memory changes as this process activates, new rules can getactivated.This can also lead to confusing situations, such as when a left-handed batter in a similarsituation is entered as data. These rules are applied in sequence and are not skipped, so morethan one state can be suggested, see below in the case of (r5) and (r6), since (r5) is detectedfirst, it suggests curve ball, but once (r6) is processed, it suggests fast ball.So in this case wedon’t fire (r7) because more than one operator was selected! No rule tells us what action to takefor multiple results!
Chunking

Upload your study docs or become a

Course Hero member to access this document

Upload your study docs or become a

Course Hero member to access this document

End of preview. Want to read all 52 pages?

Upload your study docs or become a

Course Hero member to access this document

Term
Spring
Professor
AshokK.Goel
Tags
KBAI

  • Left Quote Icon

    Student Picture

  • Left Quote Icon

    Student Picture

  • Left Quote Icon

    Student Picture