Hi, I remember an old paper by Rémi Coulom ("Computing Elo Ratings of Move Patterns in the Game of Go") where he computed "gammas" (exponentials of scores that you could feed to a softmax) for different move features, which he fit to best explain the move probabilities from real games.
Similarly, AlphaGo's paper describes how their rollout policy's weights are trained to maximize log likelihood of moves from a database. However, there is no a priori reason why imitating the probabilities observed in reference games should be optimal for this particular purpose. I thought about this for about an hour this morning, and this is what I came up with. You could make a database of positions with a label indicating the result (perhaps from real games, perhaps similarly to how AlphaGo trained their value network). Loop over the positions, run a few playouts and tweak the move probabilities by some sort of reinforcement learning, where you promote the move choices from playouts whose outcome matches the label, and you discourage the move choices from playouts whose outcome does not match the label. The point is that we would be pushing our playout policy to produce good estimates of the result of the game, which in the end is what playout policies are for. Any thoughts? Did anyone actually try something like this? Cheers, Álvaro.
_______________________________________________ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go