Lazy Incremental Learning of Control Knowledge for EfficientlyObtaining Quality Plans

• Authors: Daniel Borrajo;Manuela Veloso
• Venue: Artificial Intelligence Review - Special issue on lazy learning
• Year: 1997

General-purpose generative planners use domain-independent searchheuristics to generate solutions for problems in a variety of domains.However, in some situations these heuristics force the planner toperform inefficiently or obtain solutions of poor quality. Learningfrom experience can help to identify the particular situations forwhich the domain-independent heuristics need to be overridden. Most ofthe past learning approaches are fully deductive and eagerly acquirecorrect control knowledge from a necessarily complete domain theoryand a few examples to focus their scope. These learning strategies arehard to generalize in the case of nonlinear planning, where it isdifficult to capture correct explanations of the interactions amonggoals, multiple planning operator choices, and situational data. Inthis article, we present a lazy learning method that combines adeductive and an inductive strategy to efficiently learn controlknowledge incrementally with experience. We present hamlet, asystem we developed that learns control knowledge to improve bothsearch efficiency and the quality of the solutionsgenerated by a nonlinear planner, namely prodigy4.0. We haveidentified three lazy aspects of our approach from which we believehamlet greatly benefits: lazy explanation of successes,incremental refinement of acquired knowledge, and lazy learning tooverride only the default behavior of the problem solver. We showempirical results that support the effectiveness of this overall lazylearning approach, in terms of improving the efficiency of the problemsolver and the quality of the solutions produced.