Learning solutions to adversarial games with limited compute

How can we find optimal solution for zero-sum games (such as chess, go, or connect-four) when the state space is too large to explore exhaustively? In this work, we can learn optimal play against known classes of policies (i.e. we find limited saddle points against a restricted set of opponents). Current ongoing work of ours can efficiently find saddle-point solution to games without restricting the opponent classes. See what's currently public on Google Scholar.