TalkChess.com

There is little need to limit CPU cores for GPU engines such as Lc0 and the like. The general rule of thumb is threads should be the number of GPUs plus 1 only. More threads may show faster nps, but playing strength does not improve and could get worse. CPU RAM memory (nncache setting) is a manageab...

But, as you say, the goal is to improve strength. Innovations in Formula 1 cars aren't copyrightable if they aim to increase the car's performance rather than its aesthetics. You say tomato, I say ... For many, strong chess play is beautiful. Separately, does aesthetics have anything to do with "ar...

Machine-generation of a NN is hard to argue that a major effort of human creativity is needed to accomplish. The person arguing copyright must show that it required some personal/human creative unique effort for copyright to be possible, At least in US law. Creativity is certainly required to train...

But those are the works of human authors and creative works. They're also not derivative works of GPL'd software. But suit yourself to a double standard. Training a basic net from instructions is fairly straightforward, although there are many quite exacting steps. Carefully testing and measuring i...

Every engine has its own ecosystem of what works best. For Tinker (FWIW) my order is different: HashMove PVMove InCheck (evasions) GoodCaptures Killers (Better to do killers after captures and only store non-captures as killers) NonCaptures PoorCaptures Each move is only tried one time, so for examp...

Giraffe represents a turning point for me.

https://arxiv.org/abs/1509.01549

While not 20 years ago, in the fast-moving ML space 2015 seems pretty long ago.

I have observed something similar with Leela nets, which are not the same as NNUE nets, of course. I got significantly different results with and without draw adjudication (for example: -draw movenumber=50 movecount=6 score=10). My thinking is that what would have been considered small score differe...

I think selective or seldepth is entirely up to the engine author. Even the notion of plain old regular depth depends on the engine. There are differences between various A/B engines, and it the case of Lc0 there is no depth (it is an estimate). For seldepth in Tinker I happen to use the deepest ply...

A 5 Elo range is quite small to detect (with the 95% error) without thousands of games, unless there is a huge difference the strength of the engines. You mentioned expecting hundreds of Elo, but it might be less. Also, the time controls are very long for what is typically used with SPRT. Draw moven...

Keep in mind that "depth" with Lc0 is nothing like A/B engines.
It is an estimate based on nodes.