My program's opponents would be foolish to try mirror Go, because they do much better with regular moves!
But I can't help thinking ahead, so I thought of a trick that should help MCTS programs deal with mirror Go without biasing their play too much. In the tree search portion of MCTS, when your opponent is moving and his previous N moves have maintained symmetry, then give the first trial to a move that maintains symmetry. This does not change the progressive widening policies, so the program will have one mirror Go playout, and then follow progressive widening and RAVE rules. An alternative is to put a small (e.g. 2 RAVE wins) progressive widening bias on moves that maintain symmetry. Such techniques do not affect the asymptotic convergence properties of MCTS. So we should be OK provided that we don't overemphasize mirror Go. It occurs to me that mirror Go is often a strong strategy. Biases favoring mirror play might improve MCTS in general. Best, Brian _______________________________________________ computer-go mailing list [email protected] http://www.computer-go.org/mailman/listinfo/computer-go/
