Yet another hparameter tuner called Mainit

Discussion of anything and everything relating to chess playing software and machines.

Moderators: hgm, Rebel, chrisw

Ferdy
Posts: 4833
Joined: Sun Aug 10, 2008 3:15 pm
Location: Philippines

Yet another hparameter tuner called Mainit

Post by Ferdy »

This system is using microsoft nni - neural network intelligence toolkit. It has some tuners to use.


Sample overview dashboard. The default metric column is the Elo of the parameter values under the trial.
Image


Trial details metric, figure is Elo as objective value after every trial is completed. Every dot or trial has 64 games from a match.
Image


Parameter values selecting the top trials, that default name is Elo as objective value. Those lines represent trials, it is easy to see which trials converges to what param values. GoodEvalPruningMargin value is close to [50, 60]. MobilityWeight is close to [100, 120].
Image


Each trial is subdivided into n intermediate trials, if games per trial is 64, first will be 16 games, then second is another 16 games, after half of the games are played the average Elo will be calculated. If it is below -20, then intermediate trials 3 and 4 will not be played. The average result Elo after the full 64 games or 32 games in case of pruned trials will be reported to the optimizer.
Image

Mainit is not ready for release yet, will be released in github soon.
Ferdy
Posts: 4833
Joined: Sun Aug 10, 2008 3:15 pm
Location: Philippines

Re: Yet another hparameter tuner called Mainit

Post by Ferdy »

After 100 trials are completed, top trial is 44 with 79 elo value.

Image