Further Evolution of a Self-Learning Chess Program.

David B. Fogel, Timothy Hays, Sarah L. Hahn, James Quon

Computational Intelligence and Games · 2005 · 19 citations · 6 references

Concepts

Abstract

Previous research on the use of coevolution to improve a baseline chess program demonstrated a performance rating of 2550 against Pocket Fritz 2.0 (PF2). A series of 12 games (6 white, 6 black) was played against PF2 using the best chess program that resulted from 50 generations of variation and selection in self-play. The results yielded 9 wins, 2 losses, and 1 draw for the evolved program. This paper reports on further evolution of the best-evolved chess program, executed through 7462 generations. Results show that the outcome of this subsequent evolution was statistically significantly better than the prior player from 50 generations. A 16-game series against PF2, which plays with the rating of a high-level master, resulted in 13 wins, 0 losses, and 3 draws, yielding a performance rating of approximately 2650.

References

6