Empirical results from applying machine learning techniques to planning

Outlines an experimental machine learning implementation, called `FM', that applies both explanation based learning and similarity-based learning to AI planners. The system shell of FM contains techniques for learning application-dependent heuristics, through the experience of using a performance component (a planner) in that application. An application domain is supplied by specifying a set of action schemas, and environmental facts and rules. FM is then fed an initial state, and a sequence of tasks within this application, roughly in ascending order of complexity, which it is expected to solve. After each task has been solved, the system analyses the planning trace, allowing it to learn from experience.