Why Did You Do That? Generalizing Causal Link Explanations to Fully Observable Non-Deterministic Planning Problems

Sarath Sreedharan, Christian Muise, Subbarao Kambhampati

Abstract: The problem of designing automated agents, particularly automated planning agents that can explain their decisions has been receiving a lot of attention in recent years. The field of explainable planning or XAIP has already made a lot of progress in recent years and many of them centered around the problem of explaining decisions derived for classical planning problems. As the field progresses there is interest in tackling problems from more complex planning formalisms. However, one important aspect to keep in mind as we start focusing on such settings is that the explanatory challenges we study in the context of classical planning problems do not disappear when we move to more general settings but are just magnified. As such, when we move to these more general settings, a significant challenge before us is to see how one could generalize the well-established methods studied in the context of classical planning problems to these new settings. To provide a concrete example for this new research program we will start with causal link explanations, one of the earliest and most widely used explanations techniques used in the context of policies generated for fully observable non-deterministic planning problems. This would see us generalizing a concept that was originally developed for a specific solution concept, i.e, sequential plans, and see them applied to a very different solution concept (i.e. policies). We will develop a compilation-based method for generating generalized causal link explanations and show how as the domain is limited to deterministic cases, our method would generate causal link chains as identified by earlier works.