Detailed changes in Andscacs from 0.86 to 0.872

[cdani]

Here they are:

• * If the position is very closed, who has more space has a little more advantage.
  * Compensate something the lack of material if king is bad.
  * Penalization for weak big diagonal against king.
  * A different type of contemp. Differentiate a 0 evaluation in three different ways. 0 if the true evaluation is 0 for whatever reason. 1 if the playing side has an static evaluation > 0, and -1 on the contrary.
  * Time management:
  • - Do not stop the search too fast if the evaluation is going down.
    - Think a little more if the move done by the rival is not the one that was waiting.
  * King safety:
  • - If rooks has not been identified as attacking pieces, but they have a check, add them to the attacking number of pieces.
    - Same for knights.
    - New penalty if there are no own pawns on the sides or in front of the king.
  * Countermoves now are independent for each thread.
  * Endgame KRR(B/N) KRR(P) mainly draw.
  * Knight loses value as there are less pawns.
  * Rook wins value as there are less pawns.
  * Queen loses value as there are more rival minors.
  * Better bishop mobility if there are central mobile pawns.
  * Extra penalization for a bishop colliding with own central pawn, if this pawn is also blocked by another own pawn.
  * Little extra penalization for bishop when the rival central pawn is supported by another pawn.
  * Imbalance: If someone has passed advanced pawns, increase material advantage or compensate a little lack of material.
  * Changes of Stockfish that worked for Andscacs:
  • * http://tests.stockfishchess.org/tests/v ... 029919b260
    * http://tests.stockfishchess.org/tests/v ... 029919b27a
    * https://github.com/official-stockfish/S ... h/pull/729
  * Added singular extension: http://talkchess.com/forum/viewtopic.ph ... +extension
  * Reduce/prune less also when knight is threatening queen.
  * Little penalization for more pawn islands.
  * New prune: In low depths, and phase of the previous rival move is bad captures, and the number of root move (at depth 0) is not one of the firsts in the list, every n times this happens, prune.

Also a lot of tuning, bug solving, speed increasing,...

[brtzsnr]

* Knight loses value as there are less pawns.
* Rook wins value as there are less pawns.
* Queen loses value as there are more rival minors.

Hi, Daniel!

Thank you for sharing your progress. Do you know what is the Elo gain of each of the above features for Andscacs? In the past I found them to be noise for Zurichess at STC. I might try them again at LTC, or with the latest version of the engine.

Regards,

[sedicla]

Thanks for sharing Daniel, and good luck at TCEC!

[cdani]

* Knight loses value as there are less pawns.
* Rook wins value as there are less pawns.
* Queen loses value as there are more rival minors.

Do you know what is the Elo gain of each of the above features for Andscacs? In the past I found them to be noise for Zurichess at STC. I might try them again at LTC, or with the latest version of the engine.

Just a few elo points, but I hadn't do enough games to be able to measure the elo win seriously. Was a slight win at every time control. Sometimes this type of things are engine dependent. For example the current mg/eg pawn value can influence it, or the piece square tables, or the mobility, or whatever.

[Daniel Anulliero]

You do a very great work with Andscacs daniel
Seems you have a very good méthodology and patience , I don't have that so much!
That's why Isa's progress is very very low
Good luck for TCEC !

[voyagerOne]

Nice work Daniel!

Keep on keeping on.

-VoyagerOne

[cdani]

Thanks all! I will try to improve it more .

[cdani]

I explain better the differentiation of 1 or -1 of the draw by repetition:

* If it's draw by repetition I set a boolean.
* If the position is check I return always 0.
* If is quiescence time, I return -1 or 1 based on the full quiescence result, not on the static evaluation.
* In mate distance pruning I always return 0.
* If an evaluation is available on the tt, I return -1 or 1 based on this static evaluation, so I save a call to the evaluation function.
* And after calling the evaluation function (I have not changed the previous position of this call), I return -1 or 1 based on this value.
* I return 0 also if the static eval is 0.
* I don't do anything of this in quiescence.

This is the change that gave some win.

[Henk]

You do a very great work with Andscacs daniel
Seems you have a very good méthodology and patience , I don't have that so much!
That's why Isa's progress is very very low
Good luck for TCEC !

What I understood is that initial version of Andscacs had an ELO of about 2500. While some other engines of beginners usually start with 1300.

[mcostalba]

Here they are:

Thanks for sharing your changes, I see you have done a lot of efforts...and your efforts are paying off because you are doing a very good TCEC.

Can I ask if you are developing your engine alone and if you can share something about your test procedures (how many games, which TC, the testing hardware, etc..)

Please let me add that I am very impressed you refer to specific SF changes you found useful. This is very honest and open from you. I am sure you are not the only one to look at SF tests results, but you are the only one to state it publicly in such a transparent and detailed way: you have all my respect.