Learning to Predict by the Methods of Temporal Diierences
This article introduces a class of incremental learning procedures specialized for prediction|that is, for using past experience with an incompletely known system to predict its future behavior. Whereas conventional prediction-learning methods assign credit by means of the diierence between predicted and actual outcomes, the new methods assign credit by means of the diierence between temporally successive predictions. Although such temporal-diierence methods have been used in Samuel's checker player, Holland's bucket brigade, and the author's Adaptive Heuristic Critic, they have remained poorly understood. Here we p r o ve their convergence and optimality for special cases and relate them to supervised-learning methods. For most real-world prediction problems, temporal-diierence methods require less memory and less peak computation than conventional methodss and they produce more accurate predictions. We argue that most problems to which supervised learning is currently applied are really prediction problems of the sort to which temporal-diierence methods can be applied to advantage.