Search found 36 matches

by grahamj
Mon Feb 11, 2019 9:36 pm
Forum: Computer Chess Club: General Topics
Topic: CUDA benchmark and Elo of Leela
Replies: 3
Views: 906

Re: CUDA benchmark and Elo of Leela

There are some scaling test results here, from last August.
https://docs.google.com/spreadsheets/d/ ... 1017575449
by grahamj
Sat Feb 09, 2019 2:44 pm
Forum: Computer Chess Club: Programming and Technical Discussions
Topic: WIP, Eta - GPGPU ANN based engine, RFC
Replies: 10
Views: 2317

Re: WIP, Eta - GPGPU ANN based engine, RFC

With 'zero cost' i mean adding a qsearch to LC0 without any additional computation costs, no nps drop or similar, as thought experiment. -- Srdja I don't know enough to comment on your search algorithm. I can suggest a real experiment which might help you evaluate your q-search idea. In http://talk...
by grahamj
Sat Feb 09, 2019 8:54 am
Forum: Computer Chess Club: Programming and Technical Discussions
Topic: WIP, Eta - GPGPU ANN based engine, RFC
Replies: 10
Views: 2317

Re: WIP, Eta - GPGPU ANN based engine, RFC

I hope the Q-Search at leafnodes will do some magic. If the NN is only called to evaluate quiescent positions, it should mostly, perhaps entirely, be trained on quiescent positions. It ciould be tricky to collect suitable training data, especially If your definition of quiescent is complicated. Peo...
by grahamj
Wed Feb 06, 2019 6:43 pm
Forum: Computer Chess Club: Programming and Technical Discussions
Topic: WIP, Eta - GPGPU ANN based engine, RFC
Replies: 10
Views: 2317

Re: WIP, Eta - GPGPU ANN based engine, RFC

I hope the Q-Search at leafnodes will do some magic. If the NN is only called to evaluate quiescent positions, it should mostly, perhaps entirely, be trained on quiescent positions. It ciould be tricky to collect suitable training data, especially If your definition of quiescent is complicated. Peo...
by grahamj
Wed Feb 06, 2019 4:29 pm
Forum: Computer Chess Club: Programming and Technical Discussions
Topic: chess programming - predictions for the next 5 years
Replies: 15
Views: 3119

Re: chess programming - predictions for the next 5 years

I don't think anyone will be able to "unblackbox" neural networks in 5 years. ML researchers would love to be able to do this, but I predict very slow progress in general, and almost none in chess. I shouldn't have said "Interest in neural nets will wane...", what I really meant was "deep convolutio...
by grahamj
Wed Feb 06, 2019 4:23 pm
Forum: Computer Chess Club: Programming and Technical Discussions
Topic: WIP, Eta - GPGPU ANN based engine, RFC
Replies: 10
Views: 2317

Re: WIP, Eta - GPGPU ANN based engine, RFC

small MLP neural networks [...] * game tree in gpu memory * best node selected via score + UCT formula (visit count based) Srdja Small NNs means the evaluation will not be very good, so you need a large search tree. The UCT method requires storing the whole search tree. How will that fit in memory?
by grahamj
Wed Feb 06, 2019 7:40 am
Forum: Computer Chess Club: Programming and Technical Discussions
Topic: TensorRT Int8 results on volta
Replies: 4
Views: 1210

Re: TensorRT Int8 results on volta

Interesting.The LC0 developers tried an experiment like that, but the result in terms of ELO was disappointing. There was a suggestion that if the training games were generated using INT8 it would be better, but that wouldn't apply to supervised learning. The interactions between tensor cores, Winog...
by grahamj
Tue Feb 05, 2019 7:23 pm
Forum: Computer Chess Club: Programming and Technical Discussions
Topic: chess programming - predictions for the next 5 years
Replies: 15
Views: 3119

Re: chess programming - predictions for the next 5 years

I had no interest in chess programming until I read the AlphaZero paper. I haven't played chess for decades, and never seriously. I worked in computer vision 1995-2007, and followed the machine learning field at that time. The last ten years I have been working in mathematical biology. 15 months ago...
by grahamj
Sat Feb 02, 2019 9:50 pm
Forum: Computer Chess Club: Programming and Technical Discussions
Topic: So, how many of you are working on neural networks for chess?
Replies: 22
Views: 3985

Re: So, how many of you are working on neural networks for chess?

Its not the NN evaluation that is the problem. Its the monte-carlo tree search, and it cannot be solved with tweaks to the evaluation or the monte-carlo tree search. The solution is to stop using MCTS. It is a provably inferior tree search algorithm. The fact that its well suited to parallelization...
by grahamj
Sun Jan 20, 2019 12:06 pm
Forum: Computer Chess Club: General Topics
Topic: Why Lc0 eval (in cp) is asymmetric against AB engines?
Replies: 16
Views: 3579

Re: Why Lc0 eval (in cp) is asymmetric against AB engines?

I have found a quite different type of assymmetry. I use board positions taken randomly from various sets of games, evaluate them with both SF and LC0 using python-chess, convert to prob(win), and graph. I gave them 150ms each. The graph uses 1000 positions from 1000 games from LC0's Test20 self-pla...