Parallel evolution of game evaluation functions in ada

Abstract

This is an Ada experience report, where we conclude that Ada tasking and distributed processing facilities make it a good research tool for experimentation with algorithms that might eventually need multiple processors. We implemented a genetic algorithm in Ada to create effective computer players for Connect4. Key to our success was employing Ada tasking and ALRM Annex E Distributed computing to harness a symmetric multiproces-sor and a distributed machine with very few code changes. Easy extension of an original single-task code to multi-tasking and distributed variants-even though extension was not planned in advance-was essential to timely completion. Using either the parallel or distributed implementation, about 150 processor hours were sufficient to evolve players that neither the GNU "Four-in-a-Row" Expert player nor the author could defeat. This algorithm relies on human expertise to restrict the genetic search space. Work is in progress on a new algorithm with near-zero encoded knowledge, which will run on 220 distributed nodes within the same distributed computing framework.