Posts : 2603 Join date : 2020-11-17 Location : Netherlands
Subject: Re: TOGA IV with Rebel 14.2 NNUE evaluation Mon Mar 07, 2022 4:46 pm
Rebel 15 it will be.
Soon?
I doubt that.
Dio likes this post
matejst
Posts : 612 Join date : 2020-11-26
Subject: Re: TOGA IV with Rebel 14.2 NNUE evaluation Mon Mar 07, 2022 4:57 pm
What are your plans, Ed? Working on the code or the NN? Is there a way I could help?
Chris Whittington
Posts : 1254 Join date : 2020-11-17 Location : France
Subject: Re: TOGA IV with Rebel 14.2 NNUE evaluation Mon Mar 07, 2022 6:21 pm
matejst wrote:
What are your plans, Ed? Working on the code or the NN? Is there a way I could help?
It will very likely have a new type of net (which needs training) and some different inference code (which exists fine, but needs porting/swapping in for the prior inference code. Is one of those things where everything is in place but three different things need doing and sticking together and nothing “works” until all three work. If you’re a programmer used to working on complex structures, you’ll know that’s a highly stressy period of work. There’s no incremental, change, test, change, test cycle inbetween coffees, relax, etc. it’s just one long stress out and always difficult to summon up the determination and energy to plunge in and just do it (which is only way, basically). Programmers block.
Ah! I just had an idea …..
matejst and Dio like this post
Admin Admin
Posts : 2603 Join date : 2020-11-17 Location : Netherlands
Subject: Re: TOGA IV with Rebel 14.2 NNUE evaluation Mon Mar 07, 2022 7:02 pm
matejst wrote:
What are your plans, Ed? Working on the code or the NN? Is there a way I could help?
In the moment I am not doing much, there is a time to work and a time to enjoy the result of your work, I am extremely busy with the latter
Joking aside, next step is to use the new learner software from Chris and trying to improve the search.
matejst and Dio like this post
matejst
Posts : 612 Join date : 2020-11-26
Subject: Re: TOGA IV with Rebel 14.2 NNUE evaluation Fri Mar 11, 2022 10:36 am
Ed, just a few more remarks.
There are things I did not notice during analysis, but became clear when I watched games in matches. Anyway, I tested Toga Rebel against LGE at 1 mn 1 s in a short match. Toga Rebel's play in the middlegame was too aggressive: it would give 1, 2 pawns, a quality for the initiative, and then managed to draw with its superior play in simple positions and endgames.
This game is a good representative of Rebel's style. It is very exciting, beautiful chess. But it happens so often that it seems to hurt dramatically Rebel's results at shorter time controls.
Another thing: despite playing too risky, Rebel won the match without losing a game (+2, =10). Maitrise of simple positions and endgames is the key for good results, and there, Rebel simply shines. I noticed it already in the matches it played against Zahak and Wasp.
Chris Whittington
Posts : 1254 Join date : 2020-11-17 Location : France
Subject: Re: TOGA IV with Rebel 14.2 NNUE evaluation Fri Mar 11, 2022 5:56 pm
pohl4711 wrote:
Admin wrote:
• TOGA IV now with Rebel 14.2 NNUE evaluation. • Toga IV 1.1 is about 45 elo stronger than Rebel 14.2 and above that offers you: • Multiple threading up to 64 threads; • MultiPV up to 10 moves.
• TOGA IV 1.1 earned a rating of 3323 on the Gambit Rating List.
• Offered to you in cooperation with Chris Whittington.
Ratinglist-testrun of Toga 4.11 finished: Sorry for the mismatch in naming: Toga 4.11... Toga 4 1.11 would be correct. But, I hope, a new version comes soon by Ed Schroeder and will be named "Rebel", then this Toga-version will be deleted out of the ratinglist.
A little update on progress. I’m doing four things at once at the moment. Working on my own NNUE (training and playing with architecture), during training down-time, working like a crazy man on Chess System Tal search and Tal-ishness (coding ideas which are coming thick and fast at the moment, Elo testing and adjusting parameters) and porting some much improved NNUE code in the new Rebel, where I’m tying myself up in knots because not updating the NNUE in the right or wrong or all places in the code, it’s taking longer than expected, and fourth trying to get a good version together for the online blitz tournament tomorrow, with a book via polyglot (which I never used before) and is having the curious effect of reducing the engine to one thread only, which I don’t understand right now.
I’m not sure how Ed is getting on training a new larger net, which Rebel is going to need. Maybe in a week we’ll have something.
Mclane, TheSelfImprover, matejst and Dio like this post
matejst
Posts : 612 Join date : 2020-11-26
Subject: Re: TOGA IV with Rebel 14.2 NNUE evaluation Fri Mar 11, 2022 6:56 pm
Thanks for the news, Chris!
Mclane
Posts : 3012 Join date : 2020-11-17 Age : 57 Location : United States of Europe, Germany, Ruhr area
Subject: Re: TOGA IV with Rebel 14.2 NNUE evaluation Fri Mar 11, 2022 9:30 pm
Chris Whittington wrote:
pohl4711 wrote:
Admin wrote:
• TOGA IV now with Rebel 14.2 NNUE evaluation. • Toga IV 1.1 is about 45 elo stronger than Rebel 14.2 and above that offers you: • Multiple threading up to 64 threads; • MultiPV up to 10 moves.
• TOGA IV 1.1 earned a rating of 3323 on the Gambit Rating List.
• Offered to you in cooperation with Chris Whittington.
Ratinglist-testrun of Toga 4.11 finished: Sorry for the mismatch in naming: Toga 4.11... Toga 4 1.11 would be correct. But, I hope, a new version comes soon by Ed Schroeder and will be named "Rebel", then this Toga-version will be deleted out of the ratinglist.
A little update on progress. I’m doing four things at once at the moment. Working on my own NNUE (training and playing with architecture), during training down-time, working like a crazy man on Chess System Tal search and Tal-ishness (coding ideas which are coming thick and fast at the moment, Elo testing and adjusting parameters) and porting some much improved NNUE code in the new Rebel, where I’m tying myself up in knots because not updating the NNUE in the right or wrong or all places in the code, it’s taking longer than expected, and fourth trying to get a good version together for the online blitz tournament tomorrow, with a book via polyglot (which I never used before) and is having the curious effect of reducing the engine to one thread only, which I don’t understand right now.
I’m not sure how Ed is getting on training a new larger net, which Rebel is going to need. Maybe in a week we’ll have something.
Thanks Chris. If you need machine hardware, i have a few machines without job. Mainly 16 and 6 cores
matejst likes this post
Chris Whittington
Posts : 1254 Join date : 2020-11-17 Location : France
Subject: Re: TOGA IV with Rebel 14.2 NNUE evaluation Fri Mar 11, 2022 9:57 pm
Mclane wrote:
Chris Whittington wrote:
pohl4711 wrote:
Admin wrote:
• TOGA IV now with Rebel 14.2 NNUE evaluation. • Toga IV 1.1 is about 45 elo stronger than Rebel 14.2 and above that offers you: • Multiple threading up to 64 threads; • MultiPV up to 10 moves.
• TOGA IV 1.1 earned a rating of 3323 on the Gambit Rating List.
• Offered to you in cooperation with Chris Whittington.
Ratinglist-testrun of Toga 4.11 finished: Sorry for the mismatch in naming: Toga 4.11... Toga 4 1.11 would be correct. But, I hope, a new version comes soon by Ed Schroeder and will be named "Rebel", then this Toga-version will be deleted out of the ratinglist.
A little update on progress. I’m doing four things at once at the moment. Working on my own NNUE (training and playing with architecture), during training down-time, working like a crazy man on Chess System Tal search and Tal-ishness (coding ideas which are coming thick and fast at the moment, Elo testing and adjusting parameters) and porting some much improved NNUE code in the new Rebel, where I’m tying myself up in knots because not updating the NNUE in the right or wrong or all places in the code, it’s taking longer than expected, and fourth trying to get a good version together for the online blitz tournament tomorrow, with a book via polyglot (which I never used before) and is having the curious effect of reducing the engine to one thread only, which I don’t understand right now.
I’m not sure how Ed is getting on training a new larger net, which Rebel is going to need. Maybe in a week we’ll have something.
Thanks Chris. If you need machine hardware, i have a few machines without job. Mainly 16 and 6 cores
Ed needs more than me, I think. Self play games in enormous quantities at d9
His nets are bootstrapping off each other. When there is a good one, that net is then sued as basis for more training games and training positions. We've done that for two iterations so far, starting with Benjamin. Iteration 3 is due. 120 cores would do it in about a week.
needs PGNs with alphanumeric output. ie not cutechess. we use python software I wrote, but it needs user to download Python. and have a very fast internet, or use post.
matejst likes this post
Dio
Posts : 217 Join date : 2021-08-28
Subject: Re: TOGA IV with Rebel 14.2 NNUE evaluation Sat Mar 12, 2022 4:43 pm
I wish you good luck for today's online tournament, I hope your new progam will play interesting games.
matejst likes this post
Chris Whittington
Posts : 1254 Join date : 2020-11-17 Location : France
Subject: Re: TOGA IV with Rebel 14.2 NNUE evaluation Sat Mar 12, 2022 6:09 pm
Dio wrote:
I wish you good luck for today's online tournament, I hope your new progam will play interesting games.
Thanks. I’ll try it with a new net (only half trained right now), managed to sort out the reset of n_threads to 1 (weird problem), but not to rewrite the pondering code in time that needed redoing after some major re structural surgery on threads.
It’s got some experimental Tal concepts newly coded, I think they added something to Elo, but I’ve not had time to test much. So, either the entire pack of cards collapses under stress or it should do kind of okay. Hopefully no regressions. But. Oh, and I surrendered and added a book. I’ld prefer to play without but that means running into the same prepared lines again and again, not to mention playing against 30 book moves deep. Better would be everybody agreed an upper limit on book moves, 12 ply or something.
matejst and Dio like this post
Admin Admin
Posts : 2603 Join date : 2020-11-17 Location : Netherlands
Subject: Re: TOGA IV with Rebel 14.2 NNUE evaluation Sat Mar 12, 2022 6:54 pm
matejst wrote:
Ed, just a few more remarks.
There are things I did not notice during analysis, but became clear when I watched games in matches. Anyway, I tested Toga Rebel against LGE at 1 mn 1 s in a short match. Toga Rebel's play in the middlegame was too aggressive: it would give 1, 2 pawns, a quality for the initiative, and then managed to draw with its superior play in simple positions and endgames.
This game is a good representative of Rebel's style. It is very exciting, beautiful chess. But it happens so often that it seems to hurt dramatically Rebel's results at shorter time controls.
Another thing: despite playing too risky, Rebel won the match without losing a game (+2, =10). Maitrise of simple positions and endgames is the key for good results, and there, Rebel simply shines. I noticed it already in the matches it played against Zahak and Wasp.
Boban, you should not give me so many compliments, I get lazy from that
But your observations are correct. Within an hour going to play the monthly blitz tournament, 32 threads, whoops.
matejst likes this post
Admin Admin
Posts : 2603 Join date : 2020-11-17 Location : Netherlands
Subject: Re: TOGA IV with Rebel 14.2 NNUE evaluation Sat Mar 12, 2022 6:55 pm
Chris Whittington wrote:
Dio wrote:
I wish you good luck for today's online tournament, I hope your new progam will play interesting games.
Thanks. I’ll try it with a new net (only half trained right now), managed to sort out the reset of n_threads to 1 (weird problem), but not to rewrite the pondering code in time that needed redoing after some major re structural surgery on threads.
It’s got some experimental Tal concepts newly coded, I think they added something to Elo, but I’ve not had time to test much. So, either the entire pack of cards collapses under stress or it should do kind of okay. Hopefully no regressions. But. Oh, and I surrendered and added a book. I’ld prefer to play without but that means running into the same prepared lines again and again, not to mention playing against 30 book moves deep. Better would be everybody agreed an upper limit on book moves, 12 ply or something.
And hopefully only one SF derivative.
matejst and Dio like this post
Admin Admin
Posts : 2603 Join date : 2020-11-17 Location : Netherlands
Subject: Re: TOGA IV with Rebel 14.2 NNUE evaluation Sat Mar 12, 2022 10:12 pm
Subject: Re: TOGA IV with Rebel 14.2 NNUE evaluation Sun Mar 13, 2022 12:39 am
Ed,
There's really no reason to be disappointed. When I look at these games, the first thing I notice is the depth of the search: against CST, Rebel achieved 2-9 plies less, as an example. I also had a look at the pgn of Frank Quisinsky's KI tournament, and there too, Rebel is among the slowest 3-5 engines of 41.
There is also something Chris wrote: it is not only about the speed of the search, but about the adequacy of the search to the evaluation. Like I wrote in my previous post: Rebel favors active, aggressive and complicated positions. But in such positions its search seems to be comparatively [to other engines] even slower. Frank's stats pointed to the obvious: Rebel wins a lot of short games, but loses a lot against fast, tactical searchers [beancounters].
Rebel's evaluation is reliable -- I compared it very often with SF's evaluation at similar depths, and I tend to trust more Rebel's eval. Chris and you have obviously done a great job with NN's architectures, trainers, etc. But, if you want Elos and good results at tournaments, you have to not only speed the search, but to adapt it to an über-aggressive engine. In the world of computer chess, there's no place for Talishness, but for Nezhmetdinovness: no intuition, but calculation.
Of course, I don't have a clue how it can be done. More pruning on one hand, tactical extensions on the other? But I am sure Chris and you will figure it out, because it is also a conditio sine qua non for the development of CSTal. For my part, I just need Rebel to achieve 25 plies in the time it achieves 20 plies now -- enough for my work and pleasure.
Mclane
Posts : 3012 Join date : 2020-11-17 Age : 57 Location : United States of Europe, Germany, Ruhr area
Subject: Re: TOGA IV with Rebel 14.2 NNUE evaluation Sun Mar 13, 2022 5:08 pm
Most games were boring opening theory battles.
Chris Whittington
Posts : 1254 Join date : 2020-11-17 Location : France
Subject: Re: TOGA IV with Rebel 14.2 NNUE evaluation Sun Mar 13, 2022 6:25 pm
Mclane wrote:
Most games were boring opening theory battles.
My mistake, for first time ever mine used a large polyglot book. Is almost pointless playing with these long openings.
What I need is a small gambit book and be fast out of the opening stage.
Peter Berger
Posts : 131 Join date : 2020-11-20
Subject: Re: TOGA IV with Rebel 14.2 NNUE evaluation Sun Mar 13, 2022 6:29 pm
Mclane wrote:
Most games were boring opening theory battles.
Opening books don't matter anymore. Maybe they still do at very fast time control, but they will be made obsolete by hardware advances very soon.
People don't understand this topic very well, as they are not aware of the much more serious idea of the opening book as a liability, sth that was actually dicussed though not understood by many several years ago.
It is pretty complicated to improve on Stockfish at 3 min/move on decent hardware. This can still be done, but you can catch it only at specific points.
NNUE works way better as a general improvement on opening play.
Of course you can improve on books - and this can look pretty impressive when it works. But I am nearly 100% sure that this is a waste of time for programmers - the modern programs do a decent enough job on their own.
IMHO - YMMV Peter
matejst likes this post
matejst
Posts : 612 Join date : 2020-11-26
Subject: Re: TOGA IV with Rebel 14.2 NNUE evaluation Sun Mar 13, 2022 7:12 pm
I had a look at the CSTal-Rebel game and it was a good, instructive game. CSTal perhaps missed an opportunity to made the K,R+4p vs K, R + 3p a bit more difficult for Rebel, but it is a draw anyway. 11.Nh4 was interesting (I don't know if it was in the OB though), while 11...Ba2 seems to be a bit inferior. I would have probably prefered 0-0. After that, white had an advantage. Perhaps white should have avoided to exchange his bishop for a knight. Overall, an interesting game.
matejst
Posts : 612 Join date : 2020-11-26
Subject: Re: TOGA IV with Rebel 14.2 NNUE evaluation Sun Mar 13, 2022 7:15 pm
Peter,
you are quite right that NNs give great results in openings. And while many resulting positions are, by engines' and superGMs' standards a bit weaker, they are very, very playable at lower levels. Interactive analysis with a good engine allows a deep understanding of a particular opening.
Admin Admin
Posts : 2603 Join date : 2020-11-17 Location : Netherlands
Subject: Re: TOGA IV with Rebel 14.2 NNUE evaluation Sun Mar 13, 2022 7:28 pm
Chris Whittington wrote:
Mclane wrote:
Most games were boring opening theory battles.
My mistake, for first time ever mine used a large polyglot book. Is almost pointless playing with these long openings.
What I need is a small gambit book and be fast out of the opening stage.
Well, maybe that's my next job then.
Mclane likes this post
Sponsored content
Subject: Re: TOGA IV with Rebel 14.2 NNUE evaluation
Home 
