SAL: A Game Learning Machine Joel Paulson & Brian Lanners.

10
SAL: A Game Learning Machine Joel Paulson & Brian Joel Paulson & Brian Lanners Lanners

Transcript of SAL: A Game Learning Machine Joel Paulson & Brian Lanners.

Page 1: SAL: A Game Learning Machine Joel Paulson & Brian Lanners.

SAL: A Game Learning Machine

Joel Paulson & Brian LannersJoel Paulson & Brian Lanners

Page 2: SAL: A Game Learning Machine Joel Paulson & Brian Lanners.

Introduction

Why AI game playing?Why AI game playing? Why AI game learning?Why AI game learning? SAL (Michael Gherrity, 1993)SAL (Michael Gherrity, 1993) Search and LearningSearch and Learning

Page 3: SAL: A Game Learning Machine Joel Paulson & Brian Lanners.

Consistency Search

Basic ConceptBasic Concept Reasons for UseReasons for Use

Allows for Errors in Evaluation FunctionAllows for Errors in Evaluation Function Pathological GamesPathological Games

Page 4: SAL: A Game Learning Machine Joel Paulson & Brian Lanners.

Procedure

Consistent PositionsConsistent Positions Evaluation of a position is equal to its Evaluation of a position is equal to its

minimax valueminimax value Inconsistent PositionsInconsistent Positions

Identifying and Correcting ErrorsIdentifying and Correcting Errors

Page 5: SAL: A Game Learning Machine Joel Paulson & Brian Lanners.

General Example

If B is inconsistent, then one of the If B is inconsistent, then one of the evaluations of B, D, or E is incorrectevaluations of B, D, or E is incorrect

Page 6: SAL: A Game Learning Machine Joel Paulson & Brian Lanners.

Consistency Search in Play

Page 7: SAL: A Game Learning Machine Joel Paulson & Brian Lanners.

Organization of SAL

Game Independent KernelGame Independent KernelConsistency Search AlgorithmConsistency Search AlgorithmEvaluation FunctionsEvaluation Functions

Game Specific Move Generator Game Specific Move Generator Incorporates rules of gameIncorporates rules of gameThree Subroutines: MoveGenerator, Three Subroutines: MoveGenerator,

MakeMove, EndOfGameMakeMove, EndOfGame

Page 8: SAL: A Game Learning Machine Joel Paulson & Brian Lanners.

Features

Used as input for Neural NetworkUsed as input for Neural Network Feature Discovery ProblemFeature Discovery Problem Features in SALFeatures in SAL

Page 9: SAL: A Game Learning Machine Joel Paulson & Brian Lanners.

Neural Networks

SAL uses features for InputSAL uses features for Input Weights altered following each game using Weights altered following each game using

Temporal Difference LearningTemporal Difference Learning

Page 10: SAL: A Game Learning Machine Joel Paulson & Brian Lanners.

Performance of SAL Tic-Tac-ToeTic-Tac-Toe Connect FourConnect Four ChessChess