Fine tuning parameters

[FeikeBoomstra]

I am currently running a supposed one month long program to fine tune my parameters. The basic idea is coming from an article in the latest ICGA journal about the cross entropy method. In the end, it is really simple, it is a kind of genetic algorithm, take 100 samples, decide what are the ten best samples, calc mu and sigma and create the next generation of 100 samples.

My contribution is, that I simulated a row of swiss tournaments, to try to minimize the numbers of games to play for each series. With the simulation I found an optimal way to decrease the poule-size during the row of tournaments and two different stop criteria, to finish the series and start over with a new generation.

fyi: I currently have 29 parameters in my eval function to tune.
I just finished the first generation. And the best sample scored about 30 ratingpoints above the original.

So it is self-play with two different sets of eval-parameters. Search-depth for each move is 10 ply. Otherwise it takes too long, and my hope is, that the less your search-depth, the better the eval funtion has to be. In the extreme: if you can evaluate your next position 100% correct, you don't need any search-depth.

Hope it works out fine.

Feike

[FeikeBoomstra]

I have to re-start. Small mistake in the self-play routine.