Learning Planning Domain Descriptions in RDDL

Abstract

Recently, there is increasing interest in action model learning. However, most previous studies focused on learning effect-based action models. On the other hand, a rule-based planning domain description language was proposed in the latest planning competition. That is the Relational Dynamic Influence Diagram Language (RDDL). It uses rules to describe transitions instead of action models. In this paper, we build a system to learn planning domain descriptions in the RDDL. There are three major parts of an RDDL domain description: constraints, transitions and rewards. We first take advantage of the finite state machine analysis to identify constraints. Then, we employ the inductive learning technique to learn transitions. At last, we use regression to fix rewards. The evaluation was performed on benchmarks from planning competitions. It showed that our system can learn domain descriptions in the RDDL with low error rates. Moreover, our system is developed based on classical approaches. It implicates that the RDDL roots in previous planning languages. Therefore, more classical approaches could be useful in the RDDL domains.