Distributed game tree search on a massively parallel system

Abstract

We present our distributed αΒ-algorithm and show how αΒ- enhancements like iterative deepening, transposition tables, history tables etc. that are useful in the sequential game tree search can be applied to a distributed algorithm. The methods we describe are suitable even for large distributed systems. We describe an extension of the Young Brothers Wait Concept that we introduced to reduce the search overhead. For the first time experiments with bigger processor networks (up to 256 Transputers) show good results. We obtained a speedup of 126 running our algorithm with 256 processors.

There are mainly two reasons for this improvement. The first is that our algorithm has an inherent good load balancing, i.e. the workload using 256 processors is roughly 83% although one computation takes on the average only 300 seconds (with 256 processors).

The second reason for the good speedup achieved is the bounding of the search overhead by the extended Young Brothers Wait Concept and the efficient use of a distributed hash table. We give a cost and gain analysis of this hash table showing its superior behavior compared to other approaches.

The developed techniques have been incorparated in the distributed chess program Zugzwang, that serves as a tool for our experiments. Moreover Zugzwang participated with good results in some tournaments, for example winning the bronce medall in the 2nd Computer Games Olympiad 1990.