TalkChess.com

See here:

https://github.com/Zeta36/chess-alpha-zero

I have gotten it to work, but it is painfully slow with just one gpu.

LeelaZero is trying a more distributed approach like Fishtest, albeit for Go.

https://github.com/gcp/leela-zero

That's fast show us some games.

brianr wrote:See here:

https://github.com/Zeta36/chess-alpha-zero

I have gotten it to work, but it is painfully slow with just one gpu.

LeelaZero is trying a more distributed approach like Fishtest, albeit for Go.

https://github.com/gcp/leela-zero

That's really nice. The new Titan V from Nvidia would be a really good gpu for training. (10 times the performance of a 1080TI) but it costs 3000$

Have there been any attempts to use MCTS just for tuning of the eval weights of a "regular" chess engine?

Tuning the eval to predict the outcome of an alpha-beta search is a bit hopeless because of all the tactics that can't be encoded in a typical evaluation function. MCTS might average out the tactics sufficiently that its results can directly be used for tuning. The engine would then use alpha-beta for playing games.

syzygy wrote:Have there been any attempts to use MCTS just for tuning of the eval weights of a "regular" chess engine?

Tuning the eval to predict the outcome of an alpha-beta search is a bit hopeless because of all the tactics that can't be encoded in a typical evaluation function. MCTS might average out the tactics sufficiently that its results can directly be used for tuning. The engine would then use alpha-beta for playing games.

I dont think that would work particularly well. No matter what optimization algorithm one uses we only have a linear funciton of evaluation parameters which I doubt can encode tactics (very well).

Neural networks aren't like that and can even learn non-linear functions well.
As for tuning the "regular" evaluation parameters and dont quite understand what you mean by "using mcts to train eval parameters"

CheckersGuy wrote:As for tuning the "regular" evaluation parameters and dont quite understand what you mean by "using mcts to train eval parameters"

Run random simulations to get some sort of "averaged" estimation of the evaluation, the idea being that the tactics of the position are washed out sufficiently by the nature of mcts. Then tune the eval weights to better predict that washed out tactic-free estimation.

I don't really expect it to work, but I wouldn't expect AlphaZero Chess to work, either

Henk wrote:That's fast show us some games.

??

If this was meant for me, it would take thousands of cpu/gpu (both) hours to even just begin to train any reasonable NN. It has been estimated that AlphaGo Zero would take something like about 1,700 years on typical consumer hardware. For chess, I only got it to start running, and am still trying to understand it. There is a Connect4 game he did also that is a simpler starting point. The author is considering a distributed approach, I think.

syzygy wrote:

CheckersGuy wrote:As for tuning the "regular" evaluation parameters and dont quite understand what you mean by "using mcts to train eval parameters"

Run random simulations to get some sort of "averaged" estimation of the evaluation, the idea being that the tactics of the position are washed out sufficiently by the nature of mcts. Then tune the eval weights to better predict that washed out tactic-free estimation.

I don't really expect it to work, but I wouldn't expect AlphaZero Chess to work, either

Alpha0 NN has only tactics of depth 8 encoded in weights and it works.
MCTS should be used for playing not training.
Here is what could actually work.
Instead of alpha-beta use identical MCTS as in Alpha0.
When you reach leaf node instead of only preforming eval perform optimized mpv (of moves that wouldn't be cut by LMR) depth 8 search + QS and return score. Transform the best move score into v and ratio of move scores into probability for each move to be played all using some linear function.
If you could for example perform this mpv depth 8 search+QS in 10ms, using 64 cores machine you'd have 6400 MCT searches performed each second. That is still quite shy from 80k performed in Alpha0, but with more powerful machine you'd get there.
And despite all the hype about Alpha0's NN eval, I really doubt that it is better than depth 8 mpv search + QS of SF.

Milos wrote:

syzygy wrote:

CheckersGuy wrote:As for tuning the "regular" evaluation parameters and dont quite understand what you mean by "using mcts to train eval parameters"

Run random simulations to get some sort of "averaged" estimation of the evaluation, the idea being that the tactics of the position are washed out sufficiently by the nature of mcts. Then tune the eval weights to better predict that washed out tactic-free estimation.

I don't really expect it to work, but I wouldn't expect AlphaZero Chess to work, either

Alpha0 NN has only tactics of depth 8 encoded in weights and it works.
MCTS should be used for playing not training.
Here is what could actually work.
Instead of alpha-beta use identical MCTS as in Alpha0.
When you reach leaf node instead of only preforming eval perform optimized mpv (of moves that wouldn't be cut by LMR) depth 8 search + QS and return score. Transform the best move score into v and ratio of move scores into probability for each move to be played all using some linear function.
If you could for example perform this mpv depth 8 search+QS in 10ms, using 64 cores machine you'd have 6400 MCT searches performed each second. That is still quite shy from 80k performed in Alpha0, but with more powerful machine you'd get there.
And despite all the hype about Alpha0's NN eval, I really doubt that it is better than depth 8 mpv search + QS of SF.

I think this is more or less what Stéphane is now working on.

If I now understand the AlphaZero MCTS correctly, it is actually a sort of book-builder search. Build a tree by expanding leaf nodes and back up the results to the root, using some rule to select the next leaf node to be expanded.

If this works well with SF's alpha-beta search and evaluation, that would be very funny.

syzygy wrote:I think this is more or less what Stéphane is now working on.

If I now understand the AlphaZero MCTS correctly, it is actually a sort of book-builder search. Build a tree by expanding leaf nodes and back up the results to the root, using some rule to select the next leaf node to be expanded.

If this works well with SF's alpha-beta search and evaluation, that would be very funny.

Yes this is exactly what Stephane is working on (just found his fork a moment ago), except that he uses regular instead of mpv search.

You are right about MCTS. The rule is actually quite simple multiarmed bandit problem solution, where the nodes that have higher visit count get less explored in the future.