Grigorios Tsoumakas, Dimitris Vrakas, Nick Bassiliades, Ioannis Vlahavas
This paper describes a learning system for the automatic configuration of domain independent planning systems, based on measurable features of planning problems. The purpose of the Lazy Adaptive Multicriteria Planning (LAMP) system is to configure a planner in an optimal way, concerning two quality metrics (i.e. execution speed and plan quality), for a given problem according to user-specified preferences. The training data are produced by running the planner under consideration on a set of problems using all possible parameter configurations and recording the planning time and the plan length. When a new problem arises, LAMP extracts the values for a number of domain-expert specified problem features and uses them to identify the k nearest problems solved in the past. The system then performs a multicriteria combination of the performances of the retrieved problems according to user-specified weights that specify the relative importance of the quality metrics and selects the configuration with the best score. Experimental results show that LAMP improves the performance of the default configuration of two already well-performing planning systems in a variety of planning problems.
Keywords: Machine Learning, Planning, Instance-Based Learning, Multicriteria Combination, Feature Weighting
Citation: Grigorios Tsoumakas, Dimitris Vrakas, Nick Bassiliades, Ioannis Vlahavas: Lazy Adaptive Multicriteria Planning. In R.López de Mántaras and L.Saitta (eds.): ECAI2004, Proceedings of the 16th European Conference on Artificial Intelligence, IOS Press, Amsterdam, 2004, pp.692-696.