+ All Categories
Home > Documents > Memory and Analogy in Game-Playing Agents Jonathan Rubin & Ian Watson University of Auckland Game AI...

Memory and Analogy in Game-Playing Agents Jonathan Rubin & Ian Watson University of Auckland Game AI...

Date post: 03-Jan-2016
Category:
Upload: vivian-osborne
View: 213 times
Download: 0 times
Share this document with a friend
Popular Tags:
23
Memory and Analogy in Game-Playing Agents Jonathan Rubin & Ian Watson University of Auckland Game AI Group http://www.cs.auckland.ac.nz/ research/gameai
Transcript
Page 1: Memory and Analogy in Game-Playing Agents Jonathan Rubin & Ian Watson University of Auckland Game AI Group .

Memory and Analogy in Game-Playing Agents

Jonathan Rubin & Ian Watson

University of Auckland Game AI Grouphttp://www.cs.auckland.ac.nz/research/gameai

Page 2: Memory and Analogy in Game-Playing Agents Jonathan Rubin & Ian Watson University of Auckland Game AI Group .

Overview➲ Introduction

➲ General Game Playing

➲ Lazy Learners

➲ Memory in game-playing agents

➲ Analogical Reasoning

➲ Analogical Knowledge Transfer in GGP

➲ Conclusion

Page 3: Memory and Analogy in Game-Playing Agents Jonathan Rubin & Ian Watson University of Auckland Game AI Group .

Introduction

➲ Views and ideas about a possible approach to general game playing using memory and analogy

➲ Possible research direction

➲ Suggestions and feedback welcome

Page 4: Memory and Analogy in Game-Playing Agents Jonathan Rubin & Ian Watson University of Auckland Game AI Group .

General Game Playing

➲ Unlike specialized game players such as Deep Blue

➲ Able to play different games Accept the rules of the game

Play the game effectively without human

intervention

Page 5: Memory and Analogy in Game-Playing Agents Jonathan Rubin & Ian Watson University of Auckland Game AI Group .

Approaches to General Game Playing

➲ Partial game tree search with automated evaluation functions

➲ Approximating the minimax value by computing an exact value via simplifying abstractions of the original game

Page 6: Memory and Analogy in Game-Playing Agents Jonathan Rubin & Ian Watson University of Auckland Game AI Group .

Approaches to General Game Playing

➲ Conditional Planning (One-player games)

➲ Automatic Programming – automatic generation of programs that achieve specified objectives

Page 7: Memory and Analogy in Game-Playing Agents Jonathan Rubin & Ian Watson University of Auckland Game AI Group .

General Game PlayingOpportunities

➲ Learning

Playing multiple instances of a single game

Playing multiple games against a single player

Page 8: Memory and Analogy in Game-Playing Agents Jonathan Rubin & Ian Watson University of Auckland Game AI Group .

General Game PlayingOpportunities

➲ Identifying common lessons that can be transferred from one game instance to another

Page 9: Memory and Analogy in Game-Playing Agents Jonathan Rubin & Ian Watson University of Auckland Game AI Group .

Possible Approach toGeneral Game Playing

➲ Lazy learning approach

➲ Record a memory of experiences

➲ Analogical reasoning to generalize beyond game domains

Page 10: Memory and Analogy in Game-Playing Agents Jonathan Rubin & Ian Watson University of Auckland Game AI Group .

Lazy Learners

➲ Lazy Learners Defer processing of their inputs until they

receive requests for information (Aha,

1997)

Use local approaches

Ability to generalize well

Page 11: Memory and Analogy in Game-Playing Agents Jonathan Rubin & Ian Watson University of Auckland Game AI Group .

Memory in Games

➲ One possible definition:

Any persistent knowledge an agent has that it does not need to deduce algorithmically

Page 12: Memory and Analogy in Game-Playing Agents Jonathan Rubin & Ian Watson University of Auckland Game AI Group .

Memory-based Agents

➲ GINA – Othello (De Jong & Schultz, 1988)

➲ CHEBR – Checkers (Powell et. al., 2004)

➲ Chess (Sinclair, 1998)

➲ Casper – Poker (Rubin & Watson, 2007)

Page 13: Memory and Analogy in Game-Playing Agents Jonathan Rubin & Ian Watson University of Auckland Game AI Group .

Benefits of Memory

➲ Memory can be used to augment other approaches

Informed pruning of game tree search –

Sinclair, GINA

➲ Or, approach can be entirely based on memory alone

Casper

CHEBR

Page 14: Memory and Analogy in Game-Playing Agents Jonathan Rubin & Ian Watson University of Auckland Game AI Group .

Experience-based, Lazy learners

➲ The use of memory has been shown to be successful in a range of specialized game domains.

(Non)-Deterministic, (Im)perfect Information

➲ Lazy Learners are able to adapt well to new situations

➲ How can we extrapolate experience-based, lazy learners to handle multiple game domains?

Page 15: Memory and Analogy in Game-Playing Agents Jonathan Rubin & Ian Watson University of Auckland Game AI Group .

Analogical Knowledge Transfer

Our expertise is in PokerLet’s consider how our Poker cases could be used in an unknown game, e.g., “Monopoly”

knowledgeknowledge

Page 16: Memory and Analogy in Game-Playing Agents Jonathan Rubin & Ian Watson University of Auckland Game AI Group .

Analogical Knowledge Transfer

Poker cases have only three possible actions - Fold, Call & RaiseThese actions are useless in MonopolyBut they do provide a measure of how good or strong a Poker hand is: Fold = weak Call = OK Raise = strong

Page 17: Memory and Analogy in Game-Playing Agents Jonathan Rubin & Ian Watson University of Auckland Game AI Group .

Analogical Knowledge Transfer

A pair (two of a kind) is the most basic Poker hand

Three of a kind is stronger

Obtaining all the properties of the same colour is good in Monopoly

Page 18: Memory and Analogy in Game-Playing Agents Jonathan Rubin & Ian Watson University of Auckland Game AI Group .

Analogical Knowledge Transfer

Higher value cards in Poker are stronger than lower value cards

Higher value property is also better in Monopoly

Page 19: Memory and Analogy in Game-Playing Agents Jonathan Rubin & Ian Watson University of Auckland Game AI Group .

Analogical Knowledge Transfer

A straight in Poker is a good hand

A continuous block of properties in Monopoly increases the chances of an opponent landing on you

Page 20: Memory and Analogy in Game-Playing Agents Jonathan Rubin & Ian Watson University of Auckland Game AI Group .

Analogical Knowledge Transfer

In poker you must spend money to win money

knowledgeknowledge

Page 21: Memory and Analogy in Game-Playing Agents Jonathan Rubin & Ian Watson University of Auckland Game AI Group .

Knowledge Transfer

Superficially there is nothing in common between Poker & Monopoly

Knowledge is (in theory) transferable between the games

knowledgeknowledge

?

Page 22: Memory and Analogy in Game-Playing Agents Jonathan Rubin & Ian Watson University of Auckland Game AI Group .

Conclusion

➲ In the context of General Game playing➲ A memory-based (case-based) component

may sometimes be useful➲ Games of similar types (card, board, ...)

share concepts in common➲ Should be easier to transfer knowledge between them

➲ We believe it’s also possible to transfer knowledge between games of different types

Page 23: Memory and Analogy in Game-Playing Agents Jonathan Rubin & Ian Watson University of Auckland Game AI Group .

ThanksWe really want community feedback on this


Recommended