Evaluating aspiration window algorithm changes

[brtzsnr]

I'm evaluating a change in the my implementation of aspiration windows widening. (https://bitbucket.org/brtzsnr/zurichess ... aspiration)

So far, after 18500 games with TC 40/15s+0.05s I see no improvement.

W:5350 L:5373 D:7777 T:18500
LOS:0.4121 ELO:-0.43±3.68

Is this time control good enough for testing changes to aspiration window algorithm?

[hgm]

I think it is pointless to use games to evaluate such a change, Measuring time to depth of a few hundred representative positions should be good enough.

[brtzsnr]

I haven't had much success with using test positions to reduce search tree.

What I tried was to improve the move ordering. I took silent_but_deadly tests and ran a GA to optimize the mvvlva table I use in zurichess. In theory the GA managed to decrease the number of searched nodes by ~20%. However, overall the engine had a 25-35 ELO regression.

[hgm]

What are 'silent_but_deadly tests'?

[brtzsnr]

The tests can be found here https://chessprogramming.wikispaces.com ... but+deadly and were originally published on this forum http://www.talkchess.com/forum/viewtopi ... 3fc0f9cb4b.

[nionita]

The tests can be found here https://chessprogramming.wikispaces.com ... but+deadly and were originally published on this forum http://www.talkchess.com/forum/viewtopi ... 3fc0f9cb4b.

But those are only 130-140 positions, is it not much too less for optimising on them?

[hgm]

It seems that these are positions with a specific 'solution'. That is not representative of game positions, so you would expect tuning on them to lose Elo.