pohl4711 wrote: ↑
Sun May 27, 2018 8:31 am
mjlef wrote: ↑
Sun May 27, 2018 12:21 am
You give no evidence that a neural network is needed for Monte Carlo Tree Search
I know. Because I never said something like this. All I say is, that the interesting, risky positional of LCZero comes from its neural network, not from the MCTS, which LCZero uses only, because the neural-net calculations are so slow, that using normal AlphaBeta with crunching billions of nodes, would not work properly.
I looked into some dozend games of Komodo MCTS and find it playing like Komodo: good, solid,.positional chess, but much weaker than Komodo 12, because no real AlphaBeta search is done (seems at one level to Wasp 3). And that is, what is to be expected, when MCTS is used without a neural net.
So, Komodo MCTS has nothing in common with LCZero, but the using of MCTS, which is not only useless without a neural net, but weakening. And so, I see no reason to use Komodo MCTS.
Thats all I said. And that is all I have to say about Komodo MCTS, until it uses a neural net.
And because, I believe, that the Komodoteam knows all this like me (better like me), I think, the whole MCTS-Komodo was done to suggest parallels to AlphaZero/LC Zero, which are not existing. And I dont like that.
That has nothing to do with Komodo itself. I like Komodos positional play and its solid evaluations very much and I used only Komodo for all my work on my SALC-openings. I believe, there is no better engine for analyzing, than Komodo. Now and since the last years.
Monte Carlo Tree search is pretty well described here: https://en.wikipedia.org/wiki/Monte_Carlo_tree_search
There are many variations on it. The scheme estimates win probabilities via a number of means. A classic approach is when hitting some leve in the search tree, ply play a game with random moves from that position (called playout or rollouts), then back the result (in chess, win, loss or draw) up the tree. Each tree node contains a sum of the win/loss/draw and a count of the number of times visited. The winning percentage at each point in the tree guides it search either by exploiting what you know (favoring the best scoring move) plus exploring (favoring less visited nodes). Programs like Alpha Zero use a neural network to give win probabilities at a new node, instead of the playouts.rollouts. But it is still called Monte Carlo Tree Search since the exploitation/exploration and summing win probabilities remains the same. I think the ALpha Zero/Leela neural network is fascinating and I am anxious to see how well it can learn. But that neural network is not needed for Monte Carlo. In Komodo, we do not currently use a neural network in MCTS mode. We use a heavily modified evaluation and combined with short searches to in place of the neural network.
Larry and I have discussed this approach for years. The great performance of Alpha Zero did push us to work on it, although last tear even before Alpha Zero came out, I had talked with Larry about starting work on MCTS this year.
I do not know if Komodo's eval and search will be better or worse than Alpha Zero/Leela's neural network. I hope it will be but we have a few hundred more elo to go. But the progress since the K 12 release has been really good. We like the way it plays. And we find it coming up with very interesting moves.
I would love someone to try another experiment. Take the Leela neural network, and just write a classic search for it, using the Leela neural network in place of an eval, but include all the standard cutoffs. If easier, you could convert the win percentage back to a centipawn score, and try all the standard pruning tricks we do in programs like Stockfish. I would love to see how it performs. Neural networks can in many ways see future moves via piece cooperation, attack and defense terms. It might be a very powerful program.