evolving strategies for playing galcon

Post on 03-Jul-2015

953 Views

Category:

Entertainment & Humor

0 Downloads

Preview:

Click to see full reader

DESCRIPTION

Optimizing player behavior in a real-time strategygame using evolutionary algorithmsA. Fernández-Ares, A.M. Mora, J.J. Merelo, P. Garcí́a-Sánchez and C. Fernandes

TRANSCRIPT

1

Optimizing player behavior in a real-time strategy game using

evolutionary algorithms

A. Fernández-Ares, A.M. Mora, J.J. Merelo, P. García-Sánchez and C. Fernandes

GeNeura group: http://geneura.wordpress.com http://twitter.com/geneura

Departamento de Arquitectura y Tecnología de Computadores

University of Granada (Spain)

2

Galcon

3

Google AI Challenge

4

Game conditions

Decision time restricted to 1 second

Impossible to keep state from one step to tne next

Full knowledge of own and other's state

And Physics of the game

Swiss-style tournament

5

Baseline: GoogleBot

Planet with most ships chosen as base for attack

Planet to attack chosen according to growth rate and difference

Only one attack at the time

6

Aresbot

Hand-designed to beat GoogleBot

Similar, but

Colonies send ships to base (tithe)

If they are close to attack planet, they send attack directly

Same planet can't be attacked twice

7

But we can do better!

8

GeneBot

Let an evolutionary algorithm evolve constant and probabilities for AresBot

Use standard GA

1-elitism

BLX-α crossover

9

Several maps used to train bot

Chosen for representing different characteristics

Base planet relative position

Excentricity

If a bot is able to beat GoogleBot there, it will probably always will

10

Designing a fitness function

Important factors: number of victories and turns needed to win them

Victories are rather a constraint, so it's not multiobjective

11

Results

Genebot is able to defeat GoogleBot in the 100 maps provided by Google

Strategy completely different from AresBot

Tithe and support attack less likely, but with more ships

Growth rate fo target ship becomes more important

Many more ships sent to target planet

Beats GoogleBot in 30% less time than AresBot

12

Comparing bots

13

Did we win?

20% best: position 1454

Not bad, not great

Evolved strategy can go only as far as the underlying one.

Genetic algorithm was not really optimized

Future work.

Forget contest restrictions

Leave strategy more open

Be more “genetic”

14

Thanks for your attention - Questions?

jj@merelo.nethttp://twitter.com/jjmerelo

top related