SOAR

1
SOAR
2
The Basics

• SOAR is a theory of human cognition

• connectionist approach

• Allen Newell, one of the founders of modern
  cognitive science and artificial intelligence.

• Newells definition of intelligence

3
Introduction

• Architecture by itself does nothing.

BEHAVIOR ARCHITECTURE X CONTENT
A cognitive architecture must help produce
cognitive behavior. Soar is a theory of what
cognitive behaviors have in common.
4
(No Transcript)
5
A Professional Baseball Team

• Each position on a team could be a representation
  of an agent.
• Each agent has the over all goal of winning the
  game.
• Each has sub-goals of being successful at his
  position.
• SOAR could be used to represent this team as a
  software model.

6
Agents Make Soar Unique

• Agents are software models of real world objects.
  They are objects that react to and produce
  intelligent behavior.

• Dynamic

• Imperfect knowledge

• Prioritize the actions/ decisions

• Computational limitations

7
Agent Capabilities

• Perception- sensing the environment

• Acton- respond to the environment

• Planning- map out and decide what they will do
  before they do it.

• Learning- incorporate from their environment

• Cooperation Coordination- able to cooperate and
  coordinate with other agents using Natural
  Language Capabilities.

• Meta-Reasoning- ability to learn rationally. They
  can learn why.

8
A Scene from out Game

• Agent John Doe is a Pitcher throwing his first
  pitch of his major league career.
• He chooses to throw a curve ball.
• The batter Joe Schmoe, hits the ball, but John is
  able to catch it after it bounces between home
  plate and the pitching mound.
• John quickly throws the batter out at first base.

9
Architecture Goal Driven

• The Goal Context is the heart of the
  architecture.

• Defined by four slots and their values

• The goal The motivation, direction

• The problem space organization

• The State an internal representation
• of the situation.

• The operator the means to get from point a to
  point b.

10
Joes Goals and the State

• Joes Ultimate goal is to win the game along with
  his team
• His immediate goal is to get the batter out.
• His operators are the types of pitches that he
  can throw
• The state space includes the batter, the runners
  on base, and the current count etc.

11
Memory

• Productions are memory structures, represented
  often as if-then statements

• Constantly being matched

• Lowest level of memory

12
Productions

• There are several features that make them models
  of human memory

• They are associational in nature.

• They are independent of domain which allows for
  continuous incremental learning

• They are dynamic and cognitively impenetrable.

13
Architecture LTM

• map from the current goal context in WM to a new
  goal context.

• The mapping is from context to context, triggers
  an association

• LTM is what is true in general
• If John throws three strikes in a row then the
  batter is out.
• If John throws four balls in a row then the
  batter walks to first.

14
Architecture Working Memory

• WM arise in one of two ways external perception
  or associations

• holds the results of perception as values in
  current state.

• four kinds of objects goals, problems spaces,
  states, and operators.

• WM is what the model thinks is true in a
  particular situation
• Johns working memory tells him that he is about
  to pitch to Joe Schmoe and that he has the goal
  of getting Joe out. His choices and the current
  state are all part of the working memory. This
  changes with the batter and the status of the
  game.

15
Decision Cycle

• Moving from the general to the specific.

• the processing component that generates behavior
  out of the content that resides in the LTM and WM

• recognize-decide-act cycle

• Two Phases Elaboration and Decision

16
Architecture The Decision Cycle I, Elaboration
phase

• ALL productions that match the current state
  fire, producing new content in the working
  memory.
• For example if the batter Joe was determined to
  be left handed then a new associations would be
  made.

17
ArchitectureThe Decision Cycle II, Decision
phase

• Decision procedure interprets and suggests
  changes to the context.

• The result is either a single change, or an
  impasse
• There a limit on how much cognition can do at
  once.

18
Example Continued

• After the elaboration phase it is determined that
  John should either throw a curve ball or throw a
  fast ball based on all of the knowledge.
• Based on the possibilities the decision phase
  determines that there is not enough information
  to make a decision. There is no basis for a
  preference, an impasse has been reached.

19
Architecture Impasses

• There is an opportunity for learning.

• An impasse occurs automatically whenever there
  isnt enough knowledge.

• Independent of any domain.

• Automatically begins the creation of a new
  sub-goal context whose goal is to resolve the
  impasse.

20
Example

• Opportunity for learning.
• Past Success Rate?
• Joe recalls that in the past he is more
  successful with the curve ball.
• Throws a curve ball. Unfortunately, the ball is
  hit and John recovers the ball and throws Joe out
  at first.

21
Architecture Chunking I

• The primary learning mechanism.

• Automatically creates new associations in LTM
  whenever results are generated from an impasse.

• The new associations map relevant pre-impasse WM
  elements changes that prevent that impasse in
  future

22
Architecture Chunking II

• Chunking serves many purposes

• integrate different types of knowledge

• speed up behavior

• basis of inductive learning, analogical
  reasoning, etc.

• only architectural mechanism for changing LTM, it
  is assumed to be the basis of all types of
  learning in people.

23
Example

• Take the information, weather, time of day,
  batter, field conditions, count etc.

• Therefore the next time John may consider
  his success rate and other factors to avoid the
  impasse in the future.

• Produces preferences, a preference added that
  states if it is windy then throw less fast balls.

24
What Makes Soar Different

• ALL knowledge that is relevant is activated
  rather than just matching rules and activating
  them

• The Beliefs about the world are constantly being
  updated automatically. Not maintained by other
  rules

• Agents use Preferences. The agent can express
  knowledge about which option it prefers in the
  current situation.

• Create new and multiple states. RB systems have
  only one state

• Generates new knowledge and allows the agent to
  avoid it in the future.

25
Real World Applications

• There have been several theories developed from
  the original SOAR architecture
• NL-Soar natural language comprehension

• SCA - theory of symbolic concept

• NTD-Soar - a computational theory of perceptual,
  cognitive and motor actions performed by the NASA
  Test Director (NTD)

• IMPROV - a computational theory of how to correct
  knowledge about what actions do in the world.

26
Commercial Applications

• Soar Technologies, Inc. The goal of Soar
  Technology, Inc. is to increase the realism of
  battle simulations by developing intelligent
  automated synthetic forces.

• ExpLore Reasoning Systems, Inc. ERS builds
  intelligent software solutions for the
  mutual-fund, mortgage, credit-card, and insurance
  industries.

27
Problems With SOAR

• The programming for soars chunking can be very
  complex and difficult.

• The Einstellung Effect

• The Power Law of Learning

• There are disadvantages Soars architecture.
• large chunking may leads to incorrect knowledge

28
The Future

• Ultimately SOAR has great potential but it has
  many of the same limitations that humans have in
  learning.
• Perhaps humans arent the best model for
  intelligent behavior. Maybe there isnt a perfect
  model.
• Artificial Intelligence can and will get better.

29
References

• Lehman, Laird, Rosenbloom, A Gentle Introduction
  to Soar, an Architecture for Human Cognition
  (1993) http//ai.eecs.umich.edu/soar/main.html
• Cognitive modeling, symbolic. Wilson, Keil
  (edl.), The MIT Encyyclopedia of the Cognitive
  Sciences. Cambridge, MA MIT Press.
• Soar Technology, Soar A Comparison with
  Rule-based Systems, 2002 Soar Technology, Inc.
  http//ai.eecs.umich.edu/soar/main.html
• Soar Technology, Soar A Functional Approach to
  General Intelligence, 2002 Soar Technology, Inc.
  http//ai.eecs.umich.edu/soar/main.html
• Soar Technology Soar Along the Frontiers, 2002
  Soar Technology, Inc. http//ai.eecs.umich.edu/soa
  r/main.html
• Cognitive Theory, SOAR, Lewis, Richard L. Ohio
  State University, 1999
• http//ai.eecs.umich.edu/cogarch2/index.html.
  Cognitive Architectures