As an experienced end user of our hobby (passion) we like to hear your opinion regarding the conflicts about the fair use of NNUE and we invite you to submit your vote on one of the four questions.
We will maintain a list on this page of your vote and / or comments unless of course you want to contribute anonymously.
Your email address will be held strictly confidential and only used for verification purposes in case there is doubt on your identity.
Thanks so much in advance.
Ed Schröder July, 2021
I'm concerned about too much similarity, especially now that tuning of evaluation parameters is not considered so important by many programmers.
In fact, I am in favour of programmers creating their own, genuine network formats, like Jonathan Kreuzer, and use the HCE of their own engines to create the data.
I am not sure anyone could discover what games collection someone has used to create/trained their network, everyone uses them for books so why restrict them for NNUE?
I am definitely not interested in engines re-using the "current, public, state of the art" network, thus making them probably uniform in style. The network must be programmer's own (or team's own). Human games collection may be used either to achieve specific goals (GM style, specific skill level...) or to add a touch of human-like play to a game collection from programmer's own engine. For engine-engine fair competition, I would expect the network to "genetically" derive from the own engine lineage. Use of external "genes" (such as human games collections or other engines games collections) should be a minority holding and mandatorily made public.
Stefan Pohl (sp-cc.de)
I wouldn't describe myself as a chess engine developer - more a nnue trainer. Let me tell you a bit where I come from: Apart from nodchip, a user by the name of 'Gekkehenker' and me were the first 2 persons to ever train nnue nets for Chess (if we don't count Shogi as chess). We openly shared our data with everyone, found out better parameters and eventually beat Stockfish master.
Without collaboration it wouldn't have been possible. We helped other people and tried various net sizes and architectures. I firmly believe that the set of data you use doesn't matter too much if, and only if, you have a uniquely built neural net topology. This is the case in Stockfish, since Sopel came up with halfkaV2 specifically for Stockfish and nobody else uses it.
Of course the nets we produce are available to everybody which is great. And if you don't have a unique NN implementation and refine your data then that's great too. We want to find out new things that help the chess engine world.
Basing one's search on an engine somebody else has done and coupling that with a net trained by and for the very engine you took code from isn't exactly helping anybody in terms of ideas. The same is true for changing the net size and marketing it as a big innovation. Of course you can work on top of open source engines - nothing bad about that.
The thing that upsets me that it took years for some authors to admit that their work is a derivative of other open source engines.
IMO chess games represent raw data from which all entities should be allowed to learn.
Is this really about the pure spirit of research, or just being the BEST at any cost?