Towards the Evaluation of Action Reversibility in STRIPS Using Domain Generators

Abstract

Robustness is a major prerequisite for using AI systems in real world applications. In the context of AI planning, the reversibility of actions, i.e., the possibility to undo the effects of an action using a reverse plan, is one promising direction to achieve robust plans. Plans only made of reversible actions are resilient against goal changes during plan execution. This paper presents a naive implementation of a non-deterministic theoretical algorithm for determining action reversibility in STRIPS planning. However, evaluating action reversibility systems turns out to be a difficult challenge, as standard planning benchmarks are hardly applicable. We observed that manually crafted domains and in particular those obtained from domain generators easily contain bias. Based on an existing domain generator, we propose two slight variations that exhibit a completely different search tree characteristics. We use these domain generators to evaluate our implementation in close comparison to an existing ASP implementation and show that different generators indeed favor different implementations. Thus, a variety of domain generators is a necessary foundation for the evaluation of action reversibility systems.