13

Click here to load reader

Ants iwann jcc

Embed Size (px)

DESCRIPTION

Pres

Citation preview

Page 1: Ants iwann jcc

1

Evolving the strategies of agents for the ANTSGame

J. Carpio, P. García-Sánchez, A.M. Mora, J.J. Merelo, J. Caraballo1, F. Vaz , C. Cotta

IWANN 2013, Tenerife, 12-14 July

Page 2: Ants iwann jcc

2

Agenda

● RTS Games● Google AI Challenge● ANTS game● Bot Behaviour● Parameters to optimize● Fitness used● Experimental Setup● Results● Conclusions

Page 3: Ants iwann jcc

3

RTS Games

● Real-Time Strategy Games● Several units distributed in a playing arena that

competes for resources● Examples: StarCraft, WarCraft, Age of Empires

Page 4: Ants iwann jcc

4

Google AI Challenge

● Create AIs for RTS games: TRON (2010), Planet Wars (2011) , and ANTS (2012)

● Each submission competes with the others in the Google server

Page 5: Ants iwann jcc

5

ANTS Game

● Objective: conquer all the anthills of the enemy ● Surround enemy's ants to kill them● Collect food to create more ants● Restrictions

– Algorithms can not use previous fighting information.– Only a second per turn to move the ants colony.

Page 6: Ants iwann jcc

6

Maps

Page 7: Ants iwann jcc

7

Bot Behaviour

IF enemy hill in sight

attack the hill

ELSE IF food in sight

pick up the food

ELSE IF enemy ants in sight

attack the ants

ELSE IF non-explored zone in sight

explore the area randomly

Page 8: Ants iwann jcc

8

Parameters to optimize with the GA

● Food distance● Time remaining margin● Distances to hill/ant● Turns during left mode

Page 9: Ants iwann jcc

9

Fitness used

● Standard: using only the “score” (number of anthills conquered)

● Hierarchical: a tuple with [score, -enemy's score, number of ants, -number of enemy's ants]

Page 10: Ants iwann jcc

10

Experimental Setup

● Six different maps have been used● Each fitness evaluation is made 10 times● 64 individuals● 0,3 crossover rate● Pool of 32 best parents● After the training we compare versus the winner of

the competition and positions 165 and 993.

Page 11: Ants iwann jcc

11

Results

● Prior to optimization of the parameters, our bot ended in position 2076.

● But after optimization, it wins the bot in position 993● Number 1 and 163 were very competitive, but using

the optimized parameters with the hierarchical fitness increased the number of ants and decreased the enemy's ant

Page 12: Ants iwann jcc

12

Conclusions

● A simple agent can be optimized to win better opponents

● Hierarchical fitness increases the chances to win (adding more information)

● The strategy depends of the enemies● Future work: map analysis, online adaptation,

analysis of the enemies...

Page 13: Ants iwann jcc

13

Thanks!