Revisiting Action Failures Through the Lens of Cooperative Games

Sarath Sreedharan, David E. Smith

Keywords: Explaining Plan Failures, Game-theoretic formulation of explanations

Abstract:

This paper will revisit the question of explaining action or plan failures in the context of numeric planning problems.
As we will see, the existing methods used in classical planning for explaining plan failures fall short in this setting, and we need a novel paradigm to handle this question.
To model such explanation methods, we turn to cooperative games as a framework to capture such explanation generation problems.
Then we look at two possible methods, namely, Shapley values and Banzhaf values, to identify the role played by each action in the resulting failure.
While such methods have been previously used in explaining machine learning models, we see that their deployment in planning reveals novel challenges.
As we will see, we will need to modify how these values are calculated for the planning setting.
Finally, through a running example, we will see how explanations may be calculated using each of these methods, and how they may result in different explanations.
This points to how more work needs to be done to evaluate which of these methods may be better suited for planning settings.

Paper Type: New Short Paper

Submission Number: 5