04/01/1992 09:00 AM
Computer Science
This paper describes adaptive predictive search (APS), a learning system framework, which given little initial domain knowledge, increases its predictive abilities in complex problems domains. APS systems are termed experience-based learners because they are given as much responsibility for the learning process as possible. This framework has been applied to a number of complex domains (including chess, othello, pente, and image alignment) where the combinatorics of the state space is large and the learning process only receives reinforcement at the end of each search. The unique features of the APS framework are its pattern-weight representation of control knowledge and its integration of several learning techniques including temporal difference learning, simulated annealing, and genetic algorithms. In addition to APS, Morph, an APS chess system, is described in detail. Through training, despite little initial domain knowledge and using only one ply of search, Morph has managed several dozen draws and two wins against a traditional search based chess program stronger than most tournament players.
