Learning from Opportunities: Storing and Re-using Execution-Time Optimizations

In earlier work (Hammond 1986), we proposed a mechanism for learning from execution-time plan failure. In this paper, we suggest a corollary notion of learning from execution-time planning opportunities. We argue that both are special cases of learning from expectation failure (Schank 1982). The result of this type of learning is a set of plans for frequently occurring conjuncts of goals, indexed by the features in the world that predict their usefulness. We discuss this notion, using examples from the University of Chicago planner TRUCKER, an implementation of case-based planning in the domain of a UPS-like pickup and delivery service.