Learning Divide-and-Evolve Parameter Configurations with Racing

International audienceThe sub-optimal DAE planner implements the stochastic approach for domain-independent planning decomposition introduced in (Schoenauer, Sav´eant, and Vidal 2006; 2007). This planner optimizes either the makespan or the number of actions by generating ordered sequences of intermediate goals via a process of artificial evolution. The evolutionary part of DAE uses the Evolving Objects (EO) library, and the embedded planner it is based on is the non-optimal STRIPS planner YAHSP (Vidal 2004). For a given domain, the learning phase uses a racing procedure to choose the rates of the different variation operators used in DAE, and processes the results obtained during this process to specify the predicates that will be later used to describe the intermediate goals