Parallelizations of Game-Tree Search

This paper presents a parallelization of the ? procedure which is used in two-person game-tree searches. The main previous studies in this area are analyzed and classiied according to the three techniques of parallelization used. A new parallel algorithm founded on the exploration of the critical tree which is speciic to the sequential ? procedure and on an horizontal decomposition of game-trees is then presented. Two original techniques are proposed. The rst one is the concurrent cutoo degree k which is founded on the eeective heuristic branching factor b 1=2 or b 1? b. The second one is called the minimal generation. Empirical results obtained on a 10 processors SEQUENT BALANCE 8000 by using simulated game-trees connrm the theory that the best speedups are obtained with degree k close to the eeective heuristic branching factor b 1=2 .