TalkChess.com

I took a quick look at the content, and it seems there is extremely little machine learning in this book, except for a decision tree in Chapter 2. Mostly genetic algorithms and population-based methods. No neural network. But what they describe may be fun programming experiments.

Zeta v097 and v098 were an attempt to make use of thousands of parallel gpu- threads via an parallel Best-First-MiniMax search and classic evaluation. Very interesting, thanks. When using large neural networks, evaluation is so slow that data transfers between CPU and GPU have very little cost. To ...

I tried to reuse the current tree for the next move, but in Zeta v097/v098 I had to copy the tree back n forth between cpu/gpu, and the memory got quickly filled, so I kept it as disabled option. But in theory it should give you an boost. You store the tree in the GPU? I am very surprised. My intui...

Hi, I have also been generating an opening book automatically. My approach is simply to use MCTS to grow the book, using a long search to evaluate the leaves, and negamax backup. It is a bit similar to the drop-out expansion approach of the Othello book algorithms, but I like the MCTS approach bette...

Another generic game framework.

https://arxiv.org/abs/1908.09453
https://github.com/deepmind/open_spiel

This may be of interest to the chess variant people:
https://arxiv.org/abs/1905.05013
http://ludii.games/

Othello(Reversi)? Nice, how does it do against the big guys, e.g NTest, Edax etc? Also are you gonna release any of them? And will the whole framework be open sourced? I would be interested in a Capablanca Random Chess NN network so I keep dreaming. :D I did Othello very rapidly. It is probably not...

RL has the same problem. Weak moves of early self-play games are rapidly forgotten. In Go, the alphazero method has a very severe problem with ladders. It is similar to what you describe. A fundamental flaw of the Alpha Zero approach is that it learns only from games between strong players. When the...

Btw AG original implmentation of policy gradient did not have random sampling or dirichlet noise so there is a tendency to converge to local optima. I think they used different opponents from the pool of players at the time and only maybe discovered random sapling and noise later. The mathematical ...

Yes it makes sense to try to lower the batchsize when usinging multiple GPUs. My scorpio currently barely scales to 4 GPUs interms of nps using 128 threads per GPU (i.e. for a total 512). I have tried to optimize the performance of parallel mcts by making it completely lockless even when allocating...