How to tune search parameters ?

[Henk]

It seems to be possible to tune evaluation parameters automatically. But what about the other parameters like for instance formulas for R, LMR reduction size, futility boundaries etc. Or is it just change one parameter a bit and then play a few hundred games and see if ELO has increased.

[hgm]

more like a few thousand. But otherwise, that is the idea.

[tpetzke]

You can tune them as well

http://macechess.blogspot.de/2013/05/po ... rning.html

Thomas...

[Henk]

It says fitness function is a knockout tournament. So how many games do you play to determine the fitness of a candidate (bit vector)? [By the way genetic algorithm only finds a local optimum. ]

[mvk]

It seems to be possible to tune evaluation parameters automatically. But what about the other parameters like for instance formulas for R, LMR reduction size, futility boundaries etc. Or is it just change one parameter a bit and then play a few hundred games and see if ELO has increased.

Similar method, see last line here.

[Henk]

more like a few thousand. But otherwise, that is the idea.

I only play a few hundred games and even worse against itself. So would it be possible to go from a -> b -> c ->..... -> z where elo z < elo a
if there is a small chance that a transition from a->b is not an improvement.

[bob]

It seems to be possible to tune evaluation parameters automatically. But what about the other parameters like for instance formulas for R, LMR reduction size, futility boundaries etc. Or is it just change one parameter a bit and then play a few hundred games and see if ELO has increased.

As HGM suggested, at least a few thousand games are needed. And you can't just play ultra-fast games. Some search changes will behave differently at longer time controls and you want to be sure you don't make it better at very fast games, but worse at slower games which really count more.

[tpetzke]

So how many games do you play to determine the fitness of a candidate (bit vector)?

I increase the number of played games with each generation. In the early generations some wild candidates are generated, so it takes fewer games to find one of the stronger ones. Later the candidates are closer to each other so I play some more games (700 or so per generation). The winner is not necessarily the strongest one in that generation (far to less games for that) but it is usually a good one. That is enough for the algorithm to converge.

By the way genetic algorithm only finds a local optimum.

So do all the other optimizers presented here but unlike the others GAs have the chance to leave a local optimum again and find a better one.

[Henk]

I remember I once also used an optimizer to optimize the parameters of a genetic algorithm. Like population count, crossover probability mutation probability etc.