Fixing PDDL Domains Using Delta Debugging

Developing planning domains can be a challenging task, because existing planners provide little feedback regarding the correctness of the domain and the reasons for fail-ures in planning. In this paper we present an automated approach that can identify faulty preconditions in actions and show a working prototype implementation. Our debugger takes a faulty planning domain along with one or more problem instances as input. We use the delta debugging algorithm as the basis for our implementation, because it works by creating variations of the code in question and then testing it while essentially treating the interpreter, in our case the planner, as a black box. In our evaluation we introduce various errors into a set of selected domains taken from the International Planning Competition. We show that our debugger correctly identifies the mistakes in most cases and consistently outperforms a baseline based on brute force search.