The discussion on move evaluation via CNNs got me wondering: has anyone tried to make an evaluation function with CNNs ?
I mean, it's hard to really combine CNNs move estimator with a tree search: you still need something to tell what the best leaf is. Given the state of the art, the reflex is to use it for move ordering in the tree for MCTS. But given how strong the no-look ahead player is, it might be interesting to have a CNN generate an evaluation instead of a move, and then use alpha-beta and refinements. We probably don't want to train the final score, even if the full probability distribution is interesting; in particular, since many games end with resignation, we have missing data, and it's certainly not independant on the resignation itself. Rather take a leaf from MCTS and just predict one or zero, the evaluation function being the probability assigned to the result. Maybe a system should be found to guarantee that the move predicted by the move predictor (on 9d setting in Aja's technique) gets the highest probability of winning. (Training the boards with all alternative moves maybe ?). OK, food for thought. Jonas _______________________________________________ Computer-go mailing list Computer-go@computer-go.org http://computer-go.org/mailman/listinfo/computer-go
