Bagging strategies for learning planning policies

In this paper we describe ENSEMBLE-ROLLER, a learning-based automated planner that uses a bagging approach to enhance existing techniques for learning planning policies. Previous policy-style planning and learning systems sort state successors based on action predictions from a relational classifier. However, these learning-based planners can produce several plans of bad quality, since it is very difficult to encode in a single classifier all possible situations occurring in a planning domain. We propose to use ensembles of relational classifiers to generate more robust policies. As in other applications of machine learning, the idea of the ensembles of classifiers consists of providing accuracy for particular scenarios and diversity to cover a wide range of situations. In particular, ENSEMBLE-ROLLER learns ensembles of relational decision trees for each planning domain. The control knowledge from different sets of trees is aggregated as a single prediction or applied separately in a multiple-queue search algorithm. Experimental results show that both ways of using new policies produce on average plans of better quality.