Too much LCZero

[noobpwnftw]

I would guess that in 2-3 months we will see the beginning of the re-writing of some chapters of human opening theory at this pace.

This is not a NN exclusive thing, there are methods to develop a breadth first search tree to "solve" the openings. Just nobody is really into that, at least with a reasonable scale for chess. I don't think one can pick a particular engine and say its moves are the "right" way for the openings.

[Michel]

I wonder if the history planes could make a difference? When you feed Leela single positions the history planes are empty whereas in a game they would not be of course.

It would be interesting to see a match between Leela and a version of Leela that has its history planes disabled...

It would be interesting but will probably not happen very soon at least. We changed the network input/output planes twice now, and both times things went wrong, so the enthusiasm for those kind of changes cooled down a bit

When the project is a bit more stable with less rapid code changes, I'm sure it will be tried.

The question is if feeding Leela a position without history planes hurts her ability to analyze that position.

Therefore I was not thinking of changing the network topology. Just zeroing out the history planes before starting a new search (which would still use the history planes). If the question in the first paragraph has a positive answer then one should see a significant drop in strength in games.

[hgm]

As it seems this thread, which was about forum policy, has now been hijacked to start yet another discussion on LCZero, I will lock it, and urge people to continue this discussion in existing threads where they are on topic.