A new implementation of error analysis in game trees

Game-tree search is the core element of most attempts to make computers play games. Yet another approach is (1) to store all possible positions in a database, (2) to precompute the true values for all the positions, and (3) to extract a move with the help of a depth-one search. The databases allow computers to play optimally, in the sense that they will win every game once they have reached a winning position. Moreover, they will never lose from a drawn or won position. We find the database approach in games such as Nine-Men's-Morris and Connect-4, in endgames of chess, and in many other settings. Nevertheless, it has been observed that the database programs do not play strongly when they have to play a tournament with strong, but non-perfect human players attending. In this paper, we will present an algorithm which has the perfect knowledge of a game database as input, and whose task it is to bring an opponent without perfect knowledge into trouble. The proposed method exploits a specialised game-tree search on top of the database. More precisely, we perform an extended local look-ahead, although the true values of all positions are known. The idea is borrowed from a known theoretical error analysis in game trees. It emphasises that the combinatorial structure of the tree underlying a given position determines how difficult it is for a fallible opponent to find a correct answer. Our experiments show encouraging results.