Andscacs - New version 0.81

[cdani]

Hi!
I published a new version, 0.81. It's like 40 elo stronger, hoping that holds also on LTC.

http://andscacs.com/andscacs081.zip

* Improved pawn, knight and bishop evaluation.
* Improvements on move ordering.
* Slight improvements on king safety.
* Slight compensation of material difference when big mobility difference.
* Small speed improvements.
* Little improvements on search of moves.
* Solved various bugs (thanks to Mikko Aarnos and Chad Nunemaker!).

The tests that I'm doing are with more games (10.000-25.000 at each time control), so the changes must be more solid.

I tried to improve the evaluation function, as it plays badly a lot of positions. Anyway is far from evaluating well.

The win32 version is compatible with windows xp.

I did not do any change on lazy SMP.

I decided to publish it now because in the lasts weeks I was not able to find any serious improvements, and before trying some bigger changes I wanted to put what I have.

[Ferdy]

Wow that is very fast, I am still trying to catch up with 0.72 .
Congratz!! and thanks for releasing it.

[Graham Banks]

Hi!
I published a new version, 0.81. It's like 40 elo stronger, hoping that holds also on LTC.

http://andscacs.com/andscacs081.zip

[cdani]

Detailed changes:

* In pawn evaluation I improved isolated, doubled (reduced penalization), connected and unsupported evaluation, and added lever. Also removed a penalization for advanced weak pawn.

* If king is bad but PST of it is positive, remove PST bonus.

* Slight compensation of material difference when big mobility difference.

* On knights, improved values when it cannot be threatened by a pawn, improved value when is near the rival king. Also added a bonus when it cannot be changed by a rival minor.

* On bishops, improved value when is blocked by central pawn, improved the value when is supported by a pawn, improved PST, improved the value related to number of pawns of the same color of the bishop. Also added a bonus when it cannot be changed by a rival minor. Also added a bonus when it cannot be threatened by a pawn.

* Little knight evaluation cache, of the related values of it with pawns. It happens very often than this evaluation is the same for the next alpha_beta call. Was like 2% speed increase, not of the function, but of the whole engine.

* This of Stockfish worked nicely:
Rank based term improved approximation of pos.see() for scoring good captures.
https://github.com/official-stockfish/S ... d9a2062f32

* This of Stockfish worked nicely:
https://github.com/official-stockfish/S ... 22601a0743

* Detected and disabled some changes of previous versions that where not good, thanks to now I'm trying with more games.

* Before it was disabling some prunes when a passed pawn was moved. Now when and advanced pawn is moved.

* Increased depth of futility pruning from 4 to 6.

* Removed killer + 2 depth
http://talkchess.com/forum/viewtopic.php?t=56540

* Depth of razoring from 5 to 4.

* Improved value of king stopping passed pawn, and the value of an extra penalization for passed pawn uncontrolled.

* For each passed pawns:
if (NumberOfPwans(me)) < NumberOfPwans(rival))
eg+= eg/ 4;

[Henk]

The more games the less time you can use per game. This may have drawbacks for a deeper search needs less or different knowledge. Also what about tuning formulas for example LMR that depend on depth.

[cdani]

The more games the less time you can use per game. This may have drawbacks for the deeper the search the less (or perhaps different) knowledge it needs.

As noted some times, the improvements achieved with fast time controls tend to scale well at longer time controls, taking care of some exceptions like king safety and passed pawns. I do for example 7+0.03 and 14+0.03, and sometimes 25+0.03. Of course you need some computer power to do 10.000 games with those time controls. Anyway Andscacs arrived to like 2600 CCRL with only an quadcore I7 920, tests of maximum 1500 games, and sometimes help of other computer, of course with a lot of bad patches, but with a lot of good ones.

Also what about tuning formulas (for example LMR) that depend on depth.

The more optimized is the engine, the faster time controls you can use to optimize LMR and the like, because it reaches higher depths quicker.
Anyway, the first months I was doing may be only 500 games of long time controls like 100 seconds to try to have a minimum working LMR. You need to start with something and then slowly improve it. I'm sure I have done some hundred tests only on this search stuff. It's just a lot of work

[Frank Quisinsky]

Hi Jose,

I skip the latest Andscacs Version after your message (my News: 166). I will test the new version as soon as possible but a reorganization with my new League 2 will be run. I added your engine on my ToDo list for the moment.

Thanks for your great program.
You are very fast with improvements ... I am thinking you are working hard on it.

Best for you!
Frank

[cdani]

Hi Jose,

I skip the latest Andscacs Version after your message (my News: 166). I will test the new version as soon as possible but a reorganization with my new League 2 will be run. I added your engine on my ToDo list for the moment.

Thanks for your great program.
You are very fast with improvements ... I am thinking you are working hard on it.

Best for you!
Frank

Thanks!
I try

[Dr.Wael Deeb]

Thanks for yet another strong release....

The andscacs is approaching the gate of my +3000 Elo division with confidence in my private rating list regards,
Dr.D

[Jorge Garcia]

Congratulations Daniel, you are a monster
Keep on this tremendous pace.
Best regards from barcelona