Evaluation questions

[Henk]

I don't know about bad captures. My SEE is slow and at higher depths bad captures might turn out to be good sacrifices.
So it won't solve some tactical puzzles for it thought they were 'bad captures'

[BrokenKeyboard]

I would suggest you develop your engine step by step (testing each addition throughly in a few thousand engine vs engine matches) focusing equal attention on improving the search, eval and move ordering.

+1 to this, Having a lot of testing to figure out which updates actually help engine strength will really help guide development.
If you haven't already, I'd recommend downloading the cute-chess-cli and then running SPRT between your engine without x feature and your engine with x feature. That would allow you to determine whether or not that change really helped.

[amanjpro]

I would suggest you develop your engine step by step (testing each addition throughly in a few thousand engine vs engine matches) focusing equal attention on improving the search, eval and move ordering.

+1 to this, Having a lot of testing to figure out which updates actually help engine strength will really help guide development.
If you haven't already, I'd recommend downloading the cute-chess-cli and then running SPRT between your engine without x feature and your engine with x feature. That would allow you to determine whether or not that change really helped.

What is your setting for SPRT? I do this:

-sprt elo0=1 elo1=5 alpha=0.05 beta=0.05

But many times I see that engine_next is +1 elo points over engine_master (even if you take out the +/-) part, but the test continues... for example it is something like: 16 +/- 15, I thought this should quit the test, as the first condition is met, but I might have understood (or misunderstood really) the whole thing completely wrong

[j.t.]

What is your setting for SPRT? I do this:

-sprt elo0=1 elo1=5 alpha=0.05 beta=0.05

But many times I see that engine_next is +1 elo points over engine_master (even if you take out the +/-) part, but the test continues... for example it is something like: 16 +/- 15, I thought this should quit the test, as the first condition is met, but I might have understood (or misunderstood really) the whole thing completely wrong

I don't really understand, why one would need a different null-hypothesis that just the implicit opposite of the alternative hypothesis. If I understood correctly (I am not sure that I did!), it may work if you set elo1 also to 1 (which would mean that the null-hypthesis is that your new engine is not better than at least 1 Elo point).

[amanjpro]

What is your setting for SPRT? I do this:

-sprt elo0=1 elo1=5 alpha=0.05 beta=0.05

But many times I see that engine_next is +1 elo points over engine_master (even if you take out the +/-) part, but the test continues... for example it is something like: 16 +/- 15, I thought this should quit the test, as the first condition is met, but I might have understood (or misunderstood really) the whole thing completely wrong

I don't really understand, why one would need a different null-hypothesis that just the implicit opposite of the alternative hypothesis. If I understood correctly (I am not sure that I did!), it may work if you set elo1 also to 1 (which would mean that the null-hypthesis is that your new engine is not better than at least 1 Elo point).

Ah! I'll do test it with your suggestion, and see Thanks a lot

[RedBedHed]

Thomas and Amanj,

Thank you for these ordering schemes! I am going to try them out asap.

Thomas,

for MVV-LVA...

I think the size of the moves usually ranges from 1-100... So it would probably make sense to leverage caching when sorting...

Would you recommend insertion-sort? Or would a combination of insertion-sort and dual-pivot-quicksort be better? I know that these sorts rely on swapping pretty heavily... and dual-pivot-quicksort seems to do very well between 32-256 elements. Also, stability shouldn't matter much, right? (Would it make a difference how two "equal" moves are ordered?)

[RedBedHed]

Thomas,

Currently, I have a plain alpha-beta search that uses Zobrist keys to avoid repititions. I have implemented a piece-wise and square-wise evaluation with penalties for check and checkmate. I order moves by placing attack moves in front of passive moves.

I know it is a rough start... However, school just resumed. I won't be able to really improve things for at least a month. :/

[RedBedHed]

Staged move generation sounds very interesting, and (thanks to C++ templates) I think that I could build support into the current generator without hurting its speed too much...

[RedBedHed]

Jost, Amanj, and Het,

My stats skills are rusty lol. I'll have to see if I can fit in a quick review...

[Henk]

Best advice is to quit. But fools never listen to good advice.