Apr 24, 2007

[Other] Testing and testing

I have been playing around with a few new versions of Mediocre.

One version (which I call Mediocre v0.32full) includes all the tweaks I have tried so far, with things like; adaptive null move pruning, evaluation based LMR (it actually reduces more moves now), fractional plies, mate and move history in the board object, new extensions, restructured futility pruning, and some changes to evaluation.

I have had very mixed results with this version and until yesterday it played much weaker than v0.311.

Yesterday I tried only allowing null moves if the depth left was atleast 2 plies (which is the common way of doing it) and suddenly it started playing on par with v0.311. The results from the btm test suite is about as good as v0.311 but v0.32full searches about 1.5 deeper on average, which should be explained mainly by the greater number of moves reduced (apparently they are not all sound as of now).

I think there is a great deal of potential and there are probably a few more 'bugs' like this to work out.

No comments: