ShashChess

[Eduard]

After a few days, I played with ShashChess again on PlayChess, 2nd place tied with 1st. The best 7 today:

Friday 11 June Bullet 1+0, 19 ROUNDS 2021
(TD: Pooper Scooper#, Time: 1 min, 19 rounds)

Code: Select all

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19   
    
1 Hippo100,Stockfish 080621 bm  2460        ◊ 1/12 ◊ ½/5 ♦ ½/2 ♦ ½/3 ◊ ½/6 ◊ 1/11 ♦ ½/13 ◊ ½/10 ♦ ½/9 ◊ ½/7 ♦ ½/4 ◊ ½/3 ◊ ½/14 ♦ ½/4 ♦ ½/7 ♦ ½/9 ◊ ½/8 ◊ ½/2 ♦ ½/8    10.5/19 111.50   
2 E N,ShashChess 17.1  2400                 ◊ ½/9 ♦ 1/12 ◊ ½/1 ♦ ½/5 ◊ ½/3 ♦ ½/6 ◊ ½/4 ♦ ½/11 ◊ ½/13 ◊ ½/10 ♦ ½/8 ◊ ½/11 ♦ ½/7 ♦ ½/5 ◊ ½/8 ♦ ½/6 ◊ 1/12 ♦ ½/1 ◊ ½/3   10.5/19 105.50   
3 Taras0404,Stockfish 280521  2463          ♦ ½/13 ♦ ½/10 ◊ ½/11 ◊ ½/1 ♦ ½/2 ◊ ½/9 ♦ ½/8 ♦ ½/4 ◊ ½/5 ◊ ½/12 ◊ ½/14 ♦ ½/1 ♦ ½/6 ◊ 1/10 ◊ ½/9 ♦ ½/7 ♦ ½/4 ◊ ½/8 ♦ ½/2   10.0/19 98.00   
4 Deimus,ShashChess    2455                 ♦ ½/11 ◊ ½/13 ◊ ½/6 ♦ ½/7 ♦ ½/10 ◊ ½/8 ♦ ½/2 ◊ ½/3 ♦ ½/12 ◊ ½/14 ◊ ½/1 ♦ ½/10 ◊ ½/5 ◊ ½/1 ♦ ½/13 ♦ ½/11 ◊ ½/3 ◊ 1/9 ♦ ½/6 10.0/19 96.00   
5 Trampas,Stockfish 040621exp  2471         ◊ ½/10 ♦ ½/1 ♦ ½/13 ◊ ½/2 ♦ ½/9 ◊ ½/7 ♦ ½/6 ◊ ½/8 ♦ ½/3 ♦ ½/11 ◊ ½/12 ◊ ½/8 ♦ ½/4 ◊ ½/2 ◊ ½/6 ♦ ½/12 ◊ ½/11 ◊ ½/10 ♦ ½/7   9.5/19 95.00 159.00 
6 Maki14,Honey v13.2.1 SE  2414             ♦ ½/7 ◊ ½/9 ♦ ½/4 ◊ ½/14 ♦ ½/1 ◊ ½/2 ◊ ½/5 ◊ ½/13 ♦ ½/10 ◊ ½/8 ♦ ½/11 ♦ ½/9 ◊ ½/3 ◊ ½/7 ♦ ½/5 ◊ ½/2 ♦ ½/10 ♦ ½/12 ◊ ½/4    9.5/19 95.00 151.50 
7 SASSY_MUFFIN,Hayden 090621  2453          ◊ ½/6 ◊ ½/11  ♦ ½/10 ◊ ½/4 ♦ ½/8 ♦ ½/5 ◊ ½/9 ◊ ½/12 ◊ ½/14 ♦ ½/1 ◊ ½/13 ♦ ½/13 ◊ ½/2 ♦ ½/6 ◊ ½/1 ◊ ½/3 ♦ ½/9 ♦ ½/11 ◊ ½/5  9.5/19 95.00 149.00

1: 32 Cores
2: 11 Cores (I myself)
3: 16 Cores
4: 32 Cores
5: 16 Cores

[amchess]

Shashchess 18
https://github.com/amchess/ShashChess/releases/tag/18
On this occasion, I want to reiterate the goal of this Stockfish derivate is only to demonstrate Shashin theory can improve, at least at LTC,
original engine's strength and hard positions solving capability in a flexible way.
So, for now, we can't contribute on new nnue nets/architectures generation.
This can't be tested on the stockfish testing framework.
It also integrates our learning system on the search.
We hope our gpl code can be source of inspiration for the Stockfish community.
Already some members asked us for a collaboration and we're surely happy to contribute.
Enjoy!

[Cornfed]

Shashchess 18
https://github.com/amchess/ShashChess/releases/tag/18
On this occasion, I want to reiterate the goal of this Stockfish derivate is only to demonstrate Shashin theory can improve, at least at LTC,
original engine's strength and hard positions solving capability in a flexible way.
So, for now, we can't contribute on new nnue nets/architectures generation.
This can't be tested on the stockfish testing framework.
It also integrates our learning system on the search.
We hope our gpl code can be source of inspiration for the Stockfish community.
Already some members asked us for a collaboration and we're surely happy to contribute.
Enjoy!

As someone who has actually read Shashin's book, I am curious as to some of the testing you might have done.

For example: If one took all 40 positions in Study Chess With Tal and ran ShashChess 18 in 'Tal mode' along with, lets say, a current version of Stockfish - would you expect '18' to chose Tal's move even a little more often than SF would?

I have a feeling it would not but wondered if you had tried such a 'test'?

I have not done so, but just looked at the famous Tal - Koblencs game (20) 15.g6! (Tal) from this book. Tal's move is given as the 3rd-7th choice by ShashChess18 depending on the amount of time given.

I say I would think not because evaluations are more subjective to the human and while Tal played 15 g6 and subsequently also defended it as 'best', it may often be our, perhaps, subconscious trying to justify things often based on subsequent events. In any case, Tal/Koblencs are 'teaching ideas' here...not "best moves" which is an entirely different animal.

For example Koblencs played 15...hg6 (which Tal says "appears to be best") when clearly 15...fg6 is safer. Even if you use some algorithm to adjust the final numeric evaluations the engine comes up with, I think it would be hard to flip a final evaluation to something more 'Tal-like' if Tal's swashbuckling style would chose a move that falls outside some numeric 'range' of moves.

In this case 15.g6(!) falls well outside the numeric range of evaluations any engine (even ShashChess18) would offer when pitting it against other moves like 15. h5 or 15. Qf2 which are surely the top two choices...'coldly analytically speaking'. That is to say if the 'Tal algorithm' bumped up a particular choice up by .25 or .33 (perhaps 'Monty Carlo Style - I don't have Dragon or Komodo so I can't begin to test that) on the probability the next move or two by the opponent would be 'sub-optimal' and subsequently lead to an even larger edge for White than if 15.h5 or 15.Qf2 were chosen... 15. g6(?!!) looks like it would still fall 'outside the range' too far to make it the choice of the 'Tal Algorithm'.

[amchess]

Thanks for your post.
Tal personality means the attacking algorithm, but at the max engine's power.
So, Tal personality is a Tal's style machine!
If you want better simulate the HUMAN Tal personality, you have to use the handicap mode with Tal's elo at his time.
ShashChess has a very sophisticated handicap mode very useful overall for OTB players.
I know the book you told about, but never tried your test.
Andrea

[carldaman]

Shashchess 18
https://github.com/amchess/ShashChess/releases/tag/18
On this occasion, I want to reiterate the goal of this Stockfish derivate is only to demonstrate Shashin theory can improve, at least at LTC,
original engine's strength and hard positions solving capability in a flexible way.
So, for now, we can't contribute on new nnue nets/architectures generation.
This can't be tested on the stockfish testing framework.
It also integrates our learning system on the search.
We hope our gpl code can be source of inspiration for the Stockfish community.
Already some members asked us for a collaboration and we're surely happy to contribute.
Enjoy!

As someone who has actually read Shashin's book, I am curious as to some of the testing you might have done.

For example: If one took all 40 positions in Study Chess With Tal and ran ShashChess 18 in 'Tal mode' along with, lets say, a current version of Stockfish - would you expect '18' to chose Tal's move even a little more often than SF would?

I have a feeling it would not but wondered if you had tried such a 'test'?

I have not done so, but just looked at the famous Tal - Koblencs game (20) 15.g6! (Tal) from this book. Tal's move is given as the 3rd-7th choice by ShashChess18 depending on the amount of time given.

I say I would think not because evaluations are more subjective to the human and while Tal played 15 g6 and subsequently also defended it as 'best', it may often be our, perhaps, subconscious trying to justify things often based on subsequent events. In any case, Tal/Koblencs are 'teaching ideas' here...not "best moves" which is an entirely different animal.

For example Koblencs played 15...hg6 (which Tal says "appears to be best") when clearly 15...fg6 is safer. Even if you use some algorithm to adjust the final numeric evaluations the engine comes up with, I think it would be hard to flip a final evaluation to something more 'Tal-like' if Tal's swashbuckling style would chose a move that falls outside some numeric 'range' of moves.

In this case 15.g6(!) falls well outside the numeric range of evaluations any engine (even ShashChess18) would offer when pitting it against other moves like 15. h5 or 15. Qf2 which are surely the top two choices...'coldly analytically speaking'. That is to say if the 'Tal algorithm' bumped up a particular choice up by .25 or .33 (perhaps 'Monty Carlo Style - I don't have Dragon or Komodo so I can't begin to test that) on the probability the next move or two by the opponent would be 'sub-optimal' and subsequently lead to an even larger edge for White than if 15.h5 or 15.Qf2 were chosen... 15. g6(?!!) looks like it would still fall 'outside the range' too far to make it the choice of the 'Tal Algorithm'.

CyberNezh instantly finds 15.g6 and sticks with it as 'best'! I was curious to check it out, but it's not too surprising to see Nezh as being the most Tal-like.

I think the ShashChess Tal personality used to be more Tal-like before NNUE, which replaced the hand crafted eval.

[Graham Banks]

I think the ShashChess Tal personality used to be more Tal-like before NNUE, which replaced the hand crafted eval.

That makes sense.

[Cornfed]

I think the ShashChess Tal personality used to be more Tal-like before NNUE, which replaced the hand crafted eval.

That makes sense.

Does it? I mean, the introduction of NNUE has brought about more intuitive/sacrificial play than we ever saw with hand crafted evaluation...the very hallmark of Tal's play!

I shall have to check out this engine. I would only ever have interest in playing against an engine if it was capable of making logical but inferior moves like Tal's g6 against Koblencs.

[carldaman]

Actually, NNUE has brought in more risk-aversion, generally, along with more strength.

[amchess]

Latest 20 games LTC usual match result:
https://github.com/amchess/ShashChess/wiki/Matches

[amchess]

I think the ShashChess Tal personality used to be more Tal-like before NNUE, which replaced the hand crafted eval.

That makes sense.

Shashin theory is applied both in alpha-beta search and classical eval.
Clearly, with the nets, it's applied in alpha-beta search and very rarely in eval, because of the hybrid approach.