|
|
Post by matejst on May 14, 2022 8:34:13 GMT
Ronald Friederich published Rofchade 3.0.
When I have a look at Fred Quisinsky's KI rating list, I see:
05. Koivisto 8.6 NN : 3384.17
06. SlowChess Blitz 2.83 NN : 3374.99
07. Berserk 8.5.1 NN : 3371.63
08. rofChade 2.321 NN dev : 3368.25
It is all in the margins of error. But what's interesting is that Koivisto has a NN of 10 Mb, Slow Chess Blitz several nets for different phases of the game of 2 to 3 Mb overall, Berserk a small NN of less than one Mb, and Rofchade a NN of 24 Mb. One would think that the speed of the search is crucial -- but SCB has a way slower search than Koivisto and rofchade. Jonathan Kreuzer has put an emphasis on creating neural networks for endgames, and it looks like it's paying off. But then, Ed's and Chris' nets are, IMHO, even better (perhaps the best outside the SF nets), but there the search is a really weak point -- although I haven't not yet tested the new Rebel 15 thoroughly, in the previous iteration (Toga IV 1.1), it missed for an engine easy mates. It looks like there are no check extensions in the search. But then, what do I know? Anyway, I am perplexed.
|
|
|
|
Post by Ozymandias on May 14, 2022 10:08:36 GMT
There's way too much people trying to get into the NN action, and it's not clear they have a good idea about how they work. The SF team already did a lot of work on this front, solved a lot of problems, why not work with them? Maybe if they integrated in the framework they'd get a good idea of what works and why, and afterwards they could still do their thing, right?
|
|
|
|
Post by matejst on May 14, 2022 12:08:12 GMT
I would like to be able to combine Berserk's search with Ed's or Chris NNs, e.g. I tried Ethereal with D. Kappe's nets. While I do understand wanting originality and trying to remain original, I sometimes have problems to get some decisions, and some compatibility would be good to have. A "net runner", able to run the most common nets' formats would be an interesting project too.
I understand the problems faced by the authors: to train a big net, one needs powerful hardware and lots of time. I know it was the problem faced by the author of Zahak, and probably also by Berserk's author. Then, there is this idiotic TCEC race. Everything is streamlined just for this competition.
Seer, Koivisto have no multiPV. Bare engines. When I wrote to Connor Macmonigle that Seer crashes whenever I try to analyze an epd, he answered that it was not Seer's bug (I tried in every GUI I had). OK. They don't care if their engines are used by anybody else. Why do they publish them at all? And why open source, when they want their code not to be reused?
In retrospect, the "clone wars" from a few years ago were detrimental to computerchess. The point of the GPL/Open source in computerchess is not to share code, but to protect it. And since the conclusion of the Rybka affair was that only verbatim copying was cloning, today we don't call most of the new engines "clones", not even "derivatives", although most of them are just that -- "derivatives".
Why not leave hypocrisy behind and freely reuse code, formats? Just for the sake of experimenting, or making something ergonomic? Just for fun? Why not create a new, SF-like project? The guys working with Open Bench could easily do it -- they are all talented, and their collective project would be competitive against SF. From one point there is no logic in their behaviour.
|
|
|
|
Post by Ozymandias on May 14, 2022 15:31:39 GMT
I think they have to publish. Can a private engine participate in TCEC?
Licensing models totally lose me. I guess that's the point. Either it's private or public, all the intermediary categories only pave the road to litigation. But that's the way capitalism likes it. It should be simple: if you publish anyone can do whatever they want with it, if you're not OK with that, make it private; sell it or not. When selling it, you decide under which conditions. If the buyer agrees, he'll pay, if not, he'll pass. No lawsuits involved. At most, a breach of contract, which should be more straightforward.
|
|
