Affiliations: Dipartimento d'Ingegneria dell'Informazione, Università degli Studi di Brescia, Brescia, Italy. E-mails: gerevini@ing.unibs.it, saetti@ing.unibs.it | School of Computing and Engineering, University of Huddersfield, Huddersfield, UK. E-mail: m.vallati@hud.ac.uk
Note: [] Corresponding author: Alfonso Emilio Gerevini, Dipartimento d'Ingegneria dell'Informazione, Università degli Studi di Brescia, Via Branze 38, 25123 Brescia, Italy. E-mail: gerevini@ing.unibs.it
Abstract: The use of automatically learned knowledge for a planning domain can significantly improve the performance of a generic planner when solving a problem in this domain. In this work, we focus on the well-known SAT-based approach to planning and investigate two types of learned knowledge: macro-actions and planning horizon. Macro-actions are sequences of actions that typically occur in the solution plans, while a planning horizon of a problem is the length of a (possibly optimal) plan solving it. We propose a method that uses a machine learning tool for building a predictive model of the optimal planning horizon, and variants of the well-known planner SatPlan and solver MiniSat that can exploit macro actions and learned planning horizons to improve their performance. An experimental analysis illustrates the effectiveness of the proposed techniques demonstrating that significant speedups can be obtained.
Keywords: Planning as satisfiability, machine learning for planning
