08/01/1991 09:00 AM
Computer Science
This paper describes how a variety of machine learning methods can be combined synergistically to produce an adaptive pattern-oriented chess program. Major aspects of the following machine learning methods are used: temporal-difference learning, simulated annealing, genetic algorithms, explanation-based generalization, structured concept induction and heuristic evaluation. The need for these methods comes from the research constraints placed on the chess system. The system, \"Morph\", is limited to using just 1-ply of search, little domain knowledge and no supervised training. Thus, the system is responsible for finding a useful set of features (patterns) for evaluating states and for determining their significance (in the form of weight). To get the learning methods to cooperate effectively some design constraints normally associated with these methods need to be relaxed. The paper also argues that a benefit of a multi-strategy viewpoint is that new research ideas arise from the multiple perspectives.
