Temporal difference learning applied to a high-performance game-playing program
Jonathan Schaeffer, Markian Hlynka, Vili Jussila · International Joint Conference on Artificial Intelligence · 2001
The temporal difference (TD) learning algorithm offers the hope that the arduous task of manually tuning the evaluation function weights of game-playing programs can be automated. With one exception (TD-Gammon), TD learning has not been demonstrated to be effictive in a high-performance, world Class game-palying program. Further, there has been doubt expressed by game-program developers that learned weights could compete with the best hand-tuned weights. Chinook is the World Man-Machine tuned over 5 years. This paper shows that TD learinng is capable of competing with the best human effort.