Learning action model in AI planning based on genetic algorithm

Intelligent Planning and Machine Learning are two hot topics in AI research field. Integrated research in these two topics has gained increasing focus. Based on the assumption that no intermediate states between actions are given, this paper presents algorithms to learn action model from incomplete domain description and existent plan examples using genetic algorithm (GA). We further develop a system called AMLS-GA (Action Model Learning System Based on Genetic Algorithm) to evaluate this method. It builds a possible predicate set for each partial described action, which covers all possible predicates in precondition, add list and delete list. It encodes the action model as a hypothesis in GA search space exploiting binary coding. The whole learning process is under the standard GA framework proposed by Professor Holland in Michigan University. We define the correctness of this method as explaining more plan examples, and make comparison between the learned model and the model described by experts. Experimental results show that the algorithm can learn an action model close to expert formalism in reasonable time