An approach to solve middle game problems in Go, which was based on the probability of winning and the combinatorial game theory, was presented. Specifically, the algorithm of probability of winning was improved in this paper. With this approach, the program Midgame was implemented, which could generate and evaluate candidate moves and find the best one. Having been tested by professional Go players, Midgame is estimated at about 3 dan in terms of its middle game performance. It is accurate enough to be used by Go programs in real time and of great significance in application of computer game, artificial intelligence and games.