Experience-Based Adaptive Search (1992)

This paper introduces experience-based learning. This term refers to that type of unsupervised reinforcement learning in which almost all responsibility for the learning process is given to the system. These responsibilities include state evaluation, operator (move) selection, feature discovery and feature significance. As a learning framework experience-based learning can be applied to many problem domains (Levinson et al., 1992) . The types of problem domains considered here are restricted to complex problem domains characterized by three features. First the problem must have a formulation as a state space search. Further, reinforcement is only provided occasionally, and for many problems only at the end of a given search. Finally, the cardinality of the state space must be sufficiently large so that attempting to store all states is impractical.