intelligent online case based planning agent model for rts games conference presentation

Intelligent Online Case Based Planning Agent Model for RTS

Games

Ibrahim Fathy, Mostafa Aref, Omar Enayet, and Abdelrahman Al-Ogail

Faculty of Computer and Information Sciences

Ain-Shams University ; Cairo ; Egypt

Introduction.Problem definition.Objectives.Intelligent OLCBP Agent Model.OLCBP\RL Hybridization .Experiments and Results.Conclusion & Future Work.

Agenda

Intelligent Online Case-Based Planning Agent Model for RTS Games – ISDA’10

Online Case Based Planning is an architecture based on

CBR.

It addresses issues of plan acquisition, on-line plan

execution, interleaved planning and execution and on-line

plan adaptation.

Darmok is a previous system that applies this -without

revision- to RTS Games.

INTRODUCTION: ONLINE CASE-BASED PLANNING

Intelligent Online Case-Based Planning Agent Model for RTS Games – ISDA’10

A Reinforcement Learning Approach.

Combines temporal difference learning technique “One-

Step SARSA” with eligibility traces to learn state-action

pair values effectively.

It’s an on-policy method.

INTRODUCTION: SARSA(λ) LEARNING

Intelligent Online Case-Based Planning Agent Model for RTS Games – ISDA’10

Considered a challenging domain due to: Severe Time

Constraints – Real-Time AI – Many Objects – Imperfect

Information – Micro-Actions

INTRODUCTION: Real-Time Strategy Games

Intelligent Online Case-Based Planning Agent Model for RTS Games – ISDA’10

Research in learning and planning in real-time

strategy (RTS) games is very interesting.

The research on online case-based planning in RTS

Games does not include the capability of online

learning from experience.

The knowledge certainty remains constant, which

leads to inefficient decisions.

Problem Definition

Intelligent Online Case-Based Planning Agent Model for RTS Games – ISDA’10

Proposing an intelligent agent model based on both

online case-based planning (OLCBP) and

reinforcement learning (RL) techniques.

Increasing the certainty of the case base by learning

from experience.

Increasing both efficiency and effectiveness of the

plan decision making process.

Evaluating the model using empirical simulation on

Wargus. ( A clone of the well-known strategy game

Warcraft 2)

Objectives

Intelligent Online Case-Based Planning Agent Model for RTS Games – ISDA’10

Agent Architecture – Abstract View

Environment ( RTS Game “Wargus”)

Case Base

Offline Phase (Case

Acquisition)

Plan Expansion/Execut

ion Module

Online Case-Based

Learner

Behaviors

Goals Evaluated Case

Retrieved Case

Traces

Cases

Actions

Online Phase

Intelligent Online Case-Based Planning Agent Model for RTS Games – ISDA’10

Agent Architecture – Detailed View

Intelligent Online Case-Based Planning Agent Model for RTS Games – ISDA’10

Case RepresentationCase Goal Strategy SituationState

Shallow Features

Deep Features

BehaviorPre-Conditions

Alive-Conditions

Success-Conditions

Snippet

Learning ParametersCertainty

FactorEligibility

Prior Confidence

Intelligent Online Case-Based Planning Agent Model for RTS Games – ISDA’10

Case Representation - Example

Intelligent Online Case-Based Planning Agent Model for RTS Games – ISDA’10

Case Retrieval Since the evaluation of the case base is changed, the case

retrieval algorithm must change also. The case with the best predicted performance will be

retrieved to be executed.

Intelligent Online Case-Based Planning Agent Model for RTS Games – ISDA’10

OLCBP/RL Hybridization Algorithm

Intelligent Online Case-Based Planning Agent Model for RTS Games – ISDA’10

RL Algorithm used: SARSA(λ)

OLCBP/RL Hybridization: The Mapping.

Experiment: Case Study

Intelligent Online Case-Based Planning Agent Model for RTS Games – ISDA’10

Learning Parameter

Value used

Learning Rate (α) 0.1

Discount Rate (γ ) 0.8

Decay Rate (λ ) 0.5

Exploration Rate 0.1

Results of Attack1 & Attack2

Intelligent Online Case-Based Planning Agent Model for RTS Games – ISDA’10

Results of BuildArmy1 & BuildArmy2

Intelligent Online Case-Based Planning Agent Model for RTS Games – ISDA’10

Agent has learnt that building a smaller heavy army in that specific situation (the existence of a towers defense) is more preferable than building a larger light army. Similarly, the agent can evaluate the entire case base and learn the right choices.

Results (Cont’d)

Intelligent Online Case-Based Planning Agent Model for RTS Games – ISDA’10

Online case-based planning was hybridized with reinforcement learning in order to introduce an intelligent agent capable of planning and learning online using temporal difference with eligibility traces: Sarsa (λ) algorithm.

The empirical evaluation has shown that the proposed model –unlike Darmok System - increases the certainty of the case base by learning from experience, and hence the process of decision making for selecting more efficient, effective and successful plans.

The Paper –Conclusion

Intelligent Online Case-Based Planning Agent Model for RTS Games – ISDA’10

Implementing a prototype based on the proposed model.

Developing a strategy/case base visualization tool capable of visualizing agent’s preferred playing strategy according to its learning history. This will help in tracking the learning curve of the agent.

Finally, designing and developing a multi-agent system where agents are able to share their experiences together.

The Paper –Future Work

Intelligent Online Case-Based Planning Agent Model for RTS Games – ISDA’10

Thank You !

Questions ?

Intelligent Online Case-Based Planning Agent Model for RTS Games – ISDA’10