Studying and tackling noisy fitness in evolutionary design of game characters
Merelo, Castillo, Mora, Fernández-Ares, Esparcia-Alcázar, Cotta, Rico
We live in a noisy world
And games are even noisier
Image by James Davies
The outcome of a game is uncertain
"Eleven against eleven, and Germany wins"
How can you optimize game strategies
If you don't even know which strategy is the best
Noises have colors
Can we use the statistical properties of fitness for selection in EAs?
We will use Wilcoxon paired test for comparing chromosomes.
Testing 3 different methods
- 0-memory: reevaluate.
- Incremental temporal average: keep fitness and average.
- Wilcoxon partial order: compare using Wilcoxon test.
How WPO works every generation
For every individual
- Computes fitness again
- Does 10 comparisons with other individuals. Score +1 if better, -1 if not, 0 if not statistically significant.
- Rescale to [0,20]
Noisy experiments done this way
- Two multimodal functions, Trap and MMDP
- Additive gaussian noise with center in 0 and σ = 1,2,4
Algorithm::Evolutionary in Perl
ITA wins with low noise
WP beats it with loud noise
Different situation for MMDP
It's great to remember
Memory-based systems beat 0-memory.
WPO ~ ITA evaluation-wise, but more robust
Some way to go
- Find a more general model for noise in the fitness of games.
- Use Wilcoxon comparison in a different way: Wilcoxon Tournament
- Test in which conditions each strategy performs the best.
Thanks a lot for your attention
Use arrow keys or space bar to navigate