Andscacs - New version 0.64

[cdani]

Hi!
I published a new version of Andscacs, 064. It seems clearly stronger.

1- Some little speed optimizations and bug fixes.
2- Move ordering: At move generation, duplicate value of exchanges of pieces, and increase values of passed pawns moving to ranks 6 or 7.
3- Move generation: start at a1 for white pieces and at h8 for black pieces.
4- Cut more moves at the last alpha_beta before quiescence search.
5- Penalization for trying to win without pawns.
6- Added divide command for doing perft move by move
7- LMR in endgame: increase or decrease it depending on some conditions.
8- Don't do LMR if it's a move of pawn to the rank 7.
9- Optimized hash and solved a bug. Now it's saved also the evaluation of the position.
10- Big optimization of king safety.
11- If the evaluation of one iteration is 0, for the next one clean killers and history.
12- If at the end of quiesce there was no move analyzed, return static evaluation instead of alpha.

If someone wants more explanations, just tell me and I will expose anything with details.

May be it’s at 2800 level, but as always my tests where very limited, so let’s see what it’s able to do.

Also as always it’s a lot stronger at slow time controls, maybe even more with this version. Near all the tests have been made at 300 seconds + 0.2 seconds of increment. I mostly abandoned faster tests. The results are really different in slower time controls some times.

For example, the biggest gain seemed with the point 10. Was a clearly bad, losing 15 elo at 50+0.1, but at slower time controls won more the longer the time control was. I bet that many people had some big win attempts in his hands, and they throw it because it was losing elo at fast blitz.

For the point 11, maybe it works on Andscacs (+8 elo) because of some hidden bug on it, or maybe it’s a good idea that can be implemented in some other engines. If someone wants to try it, I hope they share their results. The idea behind it’s something like “Ok, it seems draw. Let’s try again with a new point of view”.

Direct link: http://www.andscacs.com/andscacs064.zip

[fern]

Downloaded and installed. Thanks
Fern

[cetormenter]

Hi!
11- If the evaluation of one iteration is 0, for the next one clean killers and history.

For the point 11, maybe it works on Andscacs (+8 elo) because of some hidden bug on it, or maybe it’s a good idea that can be implemented in some other engines. If someone wants to try it, I hope they share their results. The idea behind it’s something like “Ok, it seems draw. Let’s try again with a new point of view”.

Sounds like a pretty decent idea. However I tired this in Nirvana last night and maybe its simply because there are much fewer draws at the 6"+0.05" level (my typical testing conditions) but there was simply no difference between the two versions after 10000 games.

Might be interesting to submit to the StockFish testing framework as it could remove some of the early draws by repetition that are seen without changing contempt.

Anyways, congratulations on the new release it's always good to see more competition at the "lower" tier of chess engines!

[PK]

3- Move generation: start at a1 for white pieces and at h8 for black pieces.

I remember Rodent gained a bit from using fully symmetric move generator that prefers moving forward. It's funny how these little things influence move ordering even with history heuristic on the top of it.

pk

[cdani]

Downloaded and installed. Thanks

Thanks to you!

Sounds like a pretty decent idea. However I tried this in Nirvana last night and maybe its simply because there are much fewer draws at the 6"+0.05" level (my typical testing conditions) but there was simply no difference between the two versions after 10000 games.

I tried also with 6''+0.05'' and if there is an elo difference seems minimal. Very difficult to be sure. Also my tests in slow games are not with a lot of games

Might be interesting to submit to the StockFish testing framework as it could remove some of the early draws by repetition that are seen without changing contempt.

Sure. I have no idea of how to do it Maybe sometime I will spend time on those things. I’m too much centered only on Andscacs.

Anyways, congratulations on the new release it's always good to see more competition at the "lower" tier of chess engines!

Thanks also to you. Nirvanachess it's one of the standard opponents in my tests of Andscacs.

[cdani]

3- Move generation: start at a1 for white pieces and at h8 for black pieces.

I remember Rodent gained a bit from using fully symmetric move generator that prefers moving forward. It's funny how these little things influence move ordering even with history heuristic on the top of it.

pk

I will try also sooner or later to move forward first. Sure it will influence also on Andscacs

[Sylwy]

http://www.talkchess.com/forum/viewtopi ... 172#575172

Silvian R