Posts : 1518 Join date : 2020-11-17 Location : Netherlands
Subject: Re: Rebel, CS Tal Tue Jun 21, 2022 7:00 pm
Tomorrow I am starting a GRL test run of a new version from Chris.
Me, I am building a new network with a new architecture (from Chris), it's currently +60 elo in selfplay, not sure what that means against other engines.
Mclane, matejst, Damir Desevac and Dio like this post
Posts : 52 Join date : 2022-03-01 Location : Berlin
Subject: Re: Rebel, CS Tal Wed Jun 22, 2022 12:07 pm
Admin wrote:
Tomorrow I am starting a GRL test run of a new version from Chris.
Me, I am building a new network with a new architecture (from Chris), it's currently +60 elo in selfplay, not sure what that means against other engines.
So why no release of this version? +60 elo in selfplay definitly will give a nice progress in ratinglists, too. I would love doing testrun for my SPCC-ratinglist of a new version. But only, if it is released for everybody (I do not test private engines (like CEGT does (Stoofvlees-engine for example)) - one of my "golden rules"...).
Posts : 1518 Join date : 2020-11-17 Location : Netherlands
Subject: Re: Rebel, CS Tal Wed Jun 22, 2022 12:28 pm
pohl4711 wrote:
Admin wrote:
Tomorrow I am starting a GRL test run of a new version from Chris.
Me, I am building a new network with a new architecture (from Chris), it's currently +60 elo in selfplay, not sure what that means against other engines.
So why no release of this version? +60 elo in selfplay definitly will give a nice progress in ratinglists, too. I would love doing testrun for my SPCC-ratinglist of a new version. But only, if it is released for everybody (I do not test private engines (like CEGT does (Stoofvlees-engine for example)) - one of my "golden rules"...).
I will release when I am happy with a version, that means I need to polish the engine first. Steps,
1. Finish the net, will be ready today. 2. Include the search improvements in self-play and see how that goes. 3. Create an elo pool of engines compatible with the expected elo gain. 4. Test thoroughly on different time controls. 5. Release when happy.
Will probably take a week.
adminx, matejst, Damir Desevac and Dio like this post
Posts : 560 Join date : 2020-11-17 Location : France
Subject: Re: Rebel, CS Tal Wed Jun 22, 2022 3:29 pm
Admin wrote:
Tomorrow I am starting a GRL test run of a new version from Chris.
Me, I am building a new network with a new architecture (from Chris), it's currently +60 elo in selfplay, not sure what that means against other engines.
Subject: Re: Rebel, CS Tal Wed Jun 22, 2022 9:51 pm
Impressive results, indeed! Congratulation to Chris! What's catch the eye is that CSTal beat ultrafast engines AND engines with very good evaluation NNs with similar results.
And I am much more optimistic than Ed concerning Rebel -- I trust that Rebel will make a relevant Elo jump just by eliminating the bugs in the search, and the NN improvements are all bonus (although much more difficult to achieve! I also believe that Rebel will revert to the style it had initially).
Posts : 560 Join date : 2020-11-17 Location : France
Subject: Re: Rebel, CS Tal Sun Jul 03, 2022 5:45 pm
Dio wrote:
let's wait and be patient, I would also be happy to see the appearance of a new Rebel and especially CSTal, SF15 needs competition....
Whoa! I have only SF12 and SF13 in my target engine test suite. SF15 is between 120 and 200 Elo away depending on rating list.
What’s interesting though, is how we thought SF10,11,12 were unassailable monsters, yet they’ve all fallen to the stronger amateurs. LC0 likewise (although I read there’s some recent search improvement). However, I have no idea what to do to get another 200 Elo. Keep plugging on and see what happens, I guess.
Subject: Re: Rebel, CS Tal Sun Jul 03, 2022 6:01 pm
Despite some quirks with D-fend (memory), it looks very modern indeed. Lots of functionalities, good opening book, good bases, knowledge base, etc. In many aspects, at least equal and often superior to contemporary software.
In general, there are almost no software for serious work on chess today, ChessOK and Chessbase excepted. Since there is basically no competition, they simply stopped really improving their programs. I recently took advantage of the discounts at ChessOK to buy the new Aquarium and CA, and CA is practically the same as 20 years ago. It suits me, since I liked CA back then, but there are a lot of little things they could have improved.
Rebel for Windows, today more than ever, seems like a missed opportunity. I did not own Rebel 12, but I managed to buy Gandalf with the CP interface (buying chess software in the nineties and at the beginning of the millennium was almost impossible here, I had to beg my French cousins to do it for me, but it was tricky) -- and boy it was ugly! and buggy! The only good thing was the possibility to use several books at the same time.
Then, a savoir faire has been lost. I follow Pam's work at CCC, but he simply does not have a clue how to make an uncluttered and functional GUI. BSG is way better now, but it is borderline unusable still. Open source software is OK, but it leads very often to an end of diversity.
Since they really never developed new, revolutionary concepts, the SF team generally slowed down the development of new engines, and moreover, they never attempted to create a new GUI. They take the beaten path, and the beaten path always ends in a dead end. Their results are very meager for the means they have employed. Without SF, today we would have a dozen commercial engines, several GUIs, and we might have lost a hundred Elos in strength, which is insignificant.
Subject: Re: Rebel, CS Tal Sun Jul 03, 2022 6:22 pm
Chris Whittington wrote:
Dio wrote:
let's wait and be patient, I would also be happy to see the appearance of a new Rebel and especially CSTal, SF15 needs competition....
Whoa! I have only SF12 and SF13 in my target engine test suite. SF15 is between 120 and 200 Elo away depending on rating list.
What’s interesting though, is how we thought SF10,11,12 were unassailable monsters, yet they’ve all fallen to the stronger amateurs. LC0 likewise (although I read there’s some recent search improvement). However, I have no idea what to do to get another 200 Elo. Keep plugging on and see what happens, I guess.
I guess it is mostly the search. When you have a look at the SSDF rating list, you have:
which means that, for serious work, the difference between SF 12-14, Dragon, Lc0 is irrelevant, and the difference with other engines is not really important either. The search gives most Elos in bullet and blitz. Since I use engines mostly for analysis, in my experience I don't need SF. It is faster, true, but that's mostly all.