I am happy to report, that Toga 4 is not somewhat less stronger than Rebel 14.1, but somewhat stronger: https://www.sp-cc.de/
I’ld be surprised if it wasn’t. I didn’t really study in detail, just enough to get some hooks into it, but the engine source of the Toga didn’t seem very different to that of the Fruity thing, other than that someone had done the (laborious?) work of making Fruity able to run with MP = Toga. When the HCE is stripped out and replaced, that’s basically the main difference taken away. Both sets of source have been subjected by someone in the past to be more Stockfish paradigmatic (as has everything), so, running on one thread, they are likely (imo) to be fairly similar Elo-wise. Maybe Toga a little ahead because the upgraded NNUE source is maybe another 7% or so faster (nps).
matejst likes this post
matejst
Posts : 612 Join date : 2020-11-26
Subject: Re: TOGA IV with Rebel 14.1 NNUE evaluation Tue Mar 01, 2022 6:22 pm
In my experience, Toga IV is a bit faster but misses tactical shots in sharp and complicated positions and Rebel 14.1 is more reliable for use in analysis. I tried with the different Toga settings and eventually managed to find some good ones (reducing pruning a lot) to improve tactical awareness so now it seems ok. Otoh, Toga IV has multiPV that is very useful, and after reducing pruning I had to use more threads.
Overall, it would perhaps be a good idea to continue with Toga IV as the base of the new Rebel. Since I still believe that so far, the best iteration of the new Rebel is the one with Marvin (despite having a weaker net), I tend to think that for an Elo boost the search is more important than the NN. Dietrich Kappe had a net at the level of the SF nets (The White Rose), but nonetheless, his Dark Toga was several hundreds of Elos weaker. The Rebel 14.1 NN was already very good, even in late middlegames and endgames where nets are usually less "trained". A more modern search could easily give 100+ Elo points.
For my own use, I am happy with the way it is already.
Chris Whittington
Posts : 1254 Join date : 2020-11-17 Location : France
Subject: Re: TOGA IV with Rebel 14.1 NNUE evaluation Tue Mar 01, 2022 6:24 pm
Mclane wrote:
And will there be a cstal toga version too ?
Can Toga not load the Tal nets? Oh, stupid me, they’re embedded.
I should mod the code to allow a set options Weights XYZ.nnue
The reason for embedding is cutechess, and probably other GUIs in high concurrency mode.
Cutechess has no problem loading huge executables, I guess it either launches them one after the other (in effect) and waits for each one to load (on concurrency 50 with two different engines that’s 100 launches). But if the weights are loaded AFTER the exe is loaded, from HD or whatever, that’s another 100 file loads, pretty much simultaneously, and either cutechess doesn’t wait long enough, or Windows gets its knickers into a twist, or whatever, and the system breaks. Therefore, I stopped fighting the system and embedded. Then you have the disadvantage that each NNUE requires a new compile, plus people want SSE, so that doubles the compiles, and and and.
matejst likes this post
matejst
Posts : 612 Join date : 2020-11-26
Subject: Re: TOGA IV with Rebel 14.1 NNUE evaluation Tue Mar 01, 2022 6:37 pm
Sorry, Ed, I did not see your post - we were writing at the same time it seems.
Admin Admin
Posts : 2536 Join date : 2020-11-17 Location : Netherlands
Subject: Re: TOGA IV with Rebel 14.1 NNUE evaluation Tue Mar 01, 2022 7:42 pm
pohl4711 wrote:
Admin wrote:
TOGA IV - based on TOGA II 4.0 now with Rebel 14.1 NNUE evaluation.
•Toga IV is somewhat less stronger than Rebel 14.1 but offers you: • Multiple threading up to 64 threads; • MultiPV up to 10 moves.
I am happy to report, that Toga 4 is not somewhat less stronger than Rebel 14.1, but somewhat stronger: https://www.sp-cc.de/
The difference in strength between your testing and mine is not much, but I had to improve Toga IV with some Rebel goodies else Toga IV would have been 20-30 elo less. I expected the TOGA search to be a lot better.
Admin Admin
Posts : 2536 Join date : 2020-11-17 Location : Netherlands
Subject: Re: TOGA IV with Rebel 14.1 NNUE evaluation Tue Mar 01, 2022 7:45 pm
Chris Whittington wrote:
pohl4711 wrote:
Admin wrote:
TOGA IV - based on TOGA II 4.0 now with Rebel 14.1 NNUE evaluation.
•Toga IV is somewhat less stronger than Rebel 14.1 but offers you: • Multiple threading up to 64 threads; • MultiPV up to 10 moves.
I am happy to report, that Toga 4 is not somewhat less stronger than Rebel 14.1, but somewhat stronger: https://www.sp-cc.de/
I’ld be surprised if it wasn’t. I didn’t really study in detail, just enough to get some hooks into it, but the engine source of the Toga didn’t seem very different to that of the Fruity thing, other than that someone had done the (laborious?) work of making Fruity able to run with MP = Toga. When the HCE is stripped out and replaced, that’s basically the main difference taken away. Both sets of source have been subjected by someone in the past to be more Stockfish paradigmatic (as has everything), so, running on one thread, they are likely (imo) to be fairly similar Elo-wise. Maybe Toga a little ahead because the upgraded NNUE source is maybe another 7% or so faster (nps).
What bothered me in the Toga search is the pruning based on unreliable hash table scores.
matejst likes this post
Chris Whittington
Posts : 1254 Join date : 2020-11-17 Location : France
Subject: Re: TOGA IV with Rebel 14.1 NNUE evaluation Tue Mar 01, 2022 9:05 pm
Admin wrote:
pohl4711 wrote:
Admin wrote:
TOGA IV - based on TOGA II 4.0 now with Rebel 14.1 NNUE evaluation.
•Toga IV is somewhat less stronger than Rebel 14.1 but offers you: • Multiple threading up to 64 threads; • MultiPV up to 10 moves.
I am happy to report, that Toga 4 is not somewhat less stronger than Rebel 14.1, but somewhat stronger: https://www.sp-cc.de/
The difference in strength between your testing and mine is not much, but I had to improve Toga IV with some Rebel goodies else Toga IV would have been 20-30 elo less. I expected the TOGA search to be a lot better.
Oh, I forgot about those (my bad). You are correct.
matejst likes this post
matejst
Posts : 612 Join date : 2020-11-26
Subject: Re: TOGA IV with Rebel 14.1 NNUE evaluation Tue Mar 01, 2022 11:07 pm
Admin wrote:
What bothered me in the Toga search is the pruning based on unreliable hash table scores.
You were right: I switched off hash pruning and suddenly Toga's analysis is in line with Rebel's and other engines.
Peter Berger
Posts : 126 Join date : 2020-11-20
Subject: Re: TOGA IV with Rebel 14.1 NNUE evaluation Sat Mar 19, 2022 6:55 pm
A truely horrible move by Toga IV 1.1 – it played 41. Bc6. This is a position that is incredibly easy to understand for even very mediocre human players - they actually solve this by evaluation. Stockfish just does it by search – after like 3 seconds it shows the win for black.
It’s absolutely amazing how long it takes Toga to understand the position, it takes until move 46 by black.
The game was nice in other ways, too. Toga played a nice opening line, with the early development of the bishops it managed to get 10. Nbd2 in. This may not be much, but this is playing for two results, and white kept the slight edge for ever on end.
I have done a lot of games last week, but there clearly is the problem, that I can’t understand them all too well and can only judge whether I like them in general.This one I can understand and want to share,
My tech abilities are low – so if someone can transform it into someone people can look at I would be happy.
Posts : 2536 Join date : 2020-11-17 Location : Netherlands
Subject: Re: TOGA IV with Rebel 14.1 NNUE evaluation Sat Mar 19, 2022 9:03 pm
Peter Berger wrote:
A truely horrible move by Toga IV 1.1 – it played 41. Bc6.
Here SF 14.1 needs 62 seconds and 37 ply to see that Bc6 is bad. And Toga isn't a search monster. I tried the number 5 from the ccrl rating list and after 10 minutes and depth 34 it still has Bc6 as best move. Meaning, you have found yourself a nice positional test position.
3n1k2/p2B4/Pp1P2pp/1P1ppp2/7P/4P1P1/5PK1/8 w am Bc6;
TheSelfImprover
Posts : 3079 Join date : 2020-11-18
Subject: Re: TOGA IV with Rebel 14.1 NNUE evaluation Sat Mar 19, 2022 9:09 pm
Eelco wrote:
Hi Peter,
showing the game here is easy enough, you just have to include the game score between square brackets so (pgn)(/pgn) but then with square brackets:
This forum's analysis board (click the E8 square in the quoted text above) also wants to play 41.Bc6
Mclane
Posts : 3009 Join date : 2020-11-17 Age : 57 Location : United States of Europe, Germany, Ruhr area
Subject: Re: TOGA IV with Rebel 14.1 NNUE evaluation Sat Mar 19, 2022 10:16 pm
So nothing to test ?!
Peter Berger
Posts : 126 Join date : 2020-11-20
Subject: Re: TOGA IV with Rebel 14.1 NNUE evaluation Mon Mar 21, 2022 2:34 pm
Admin wrote:
Peter Berger wrote:
A truely horrible move by Toga IV 1.1 – it played 41. Bc6.
Here SF 14.1 needs 62 seconds and 37 ply to see that Bc6 is bad. And Toga isn't a search monster. I tried the number 5 from the ccrl rating list and after 10 minutes and depth 34 it still has Bc6 as best move. Meaning, you have found yourself a nice positional test position.
3n1k2/p2B4/Pp1P2pp/1P1ppp2/7P/4P1P1/5PK1/8 w am Bc6;
This means that it is actually 41. f4! best and only move, which might make it a pretty decent testposition indeed
Admin likes this post
Sponsored content
Subject: Re: TOGA IV with Rebel 14.1 NNUE evaluation