Controlling bots in UT2004 using genetic algorithms (WCCI 2010)

1
CONTROLLING BOTS IN A FIRST-PERSON SHOOTER GAME USING GENETIC ALGORITHMS Anna I Esparcia-Alcázar 1 , Anaís Martínez-García 1 , Antonio M. Mora 2 , J.J. Merelo 2 , Pablo García-Sánchez 2 1 Instituto Tecnológico de Informática - Universidad Politécnica de Valencia, Spain 2 Dept. of Architecture and Computer Technology - University of Granada, Spain THE ENGINE Unreal Tournament 2004 • A First-Person Shooter (FPS) game • Bots controlled by very advanced AI • Use of open language and environment • Several kinds of games → we focus on DeathMatch AI IN UT2004 Each bot is controlled by a Fuzzy Finite State Machine (FFSM), for which Transitions depend on: • Bot personality, e.g. aggressiveness. • Environment situation, such as distance to enemy. • Bot condition, for example, health States represent situations and objectives. For instance, Attacking or Hunting EVOLVING BOTS We use a GA to evolve the values that influence transitions in the FFSM A chromosome consists of 56 genes. For evaluation, the bot controlled by each chromosome is given 7 lives. We test four fitness functions: RESULTS & CONCLUSIONS • Great differences between evolution stage and at play time. • Incorporating the lifespan into the fitness provides better results. • Aiming for only one objective does not produce interesting behaviours. • MultiBots develop a picking behaviour, which they are not aimed for. NoHNESProject (ref. TIN2007-68083-C02) Partially funded by: OBJECTIVE To evolve interesting bot behaviours using simple fitness functions KillerBot maximise kills PickerBot maximise items collected MultiBot combination of kills and lifespan (*) (*) using NSGAII SurvivorBot maximise lifespan FUTURE WORK • Develop teams of bots, with cooperating behaviours. • Imitate humans!

description

The aim of this work is to use Genetic Algorithms to evolve interesting bot behaviours in Unreal Tournament 2004 using simple fitness functions This poster was presented at the IEEE-WCCI 2010 congress, in Barcelona, Spain

Transcript of Controlling bots in UT2004 using genetic algorithms (WCCI 2010)

Page 1: Controlling bots in UT2004 using genetic algorithms (WCCI 2010)

CONTROLLING BOTS IN A FIRST-PERSON SHOOTER GAME USING GENETIC ALGORITHMS

Anna I Esparcia-Alcázar1, Anaís Martínez-García1, Antonio M. Mora2, J.J. Merelo2, Pablo García-Sánchez2

1Instituto Tecnológico de Informática - Universidad Politécnica de Valencia, Spain2Dept. of Architecture and Computer Technology - University of Granada, Spain

THE ENGINE

Unreal Tournament 2004

• A First-Person Shooter (FPS) game

• Bots controlled by very advanced AI

• Use of open language and environment

• Several kinds of games → we focus on DeathMatch

AI IN UT2004

Each bot is controlled by a Fuzzy Finite State Machine (FFSM), for which

• Transitions depend on:

• Bot personality, e.g. aggressiveness.

• Environment situation, such as distance to enemy.

• Bot condition, for example, health

• States represent situations and objectives. For instance, Attacking or Hunting

EVOLVING BOTS

We use a GA to evolve the values that influence transitions in the FFSM

A chromosome consists of 56 genes.

For evaluation, the bot controlled by each chromosome is given 7 lives.

We test four fitness functions:

RESULTS & CONCLUSIONS

• Great differences between evolution stage and at play time.

• Incorporating the lifespan into the fitness provides better results.

• Aiming for only one objective does not produce interesting behaviours.

• MultiBots develop a picking behaviour, which they are not aimed for.

NoHNESProject(ref. TIN2007-68083-C02)

Partially funded by:

OBJECTIVE To evolve interesting bot behaviours using simple fitness functions

KillerBot maximise kills

PickerBotmaximise items collected

MultiBotcombination of kills

and lifespan (*)

(*) using NSGAII

SurvivorBot maximise lifespan

FUTURE WORK

• Develop teams of bots, with cooperating behaviours.

• Imitate humans!