A General Logic-based Approach for Explanation Generation

Anonymous

Keywords: Knowledge Representation and Reasoning/Automated Reasoning and Theorem Proving, Explainable AI

TL;DR: A general framework for explanation generation using Logic.

Abstract: In an explanation generation problem, an agent needs to identify and explain the reasons for its decisions to another agent. Existing work in this area is mostly confined to planning-based systems that use automated planning approaches to solve the problem. In this paper, we approach this problem from a new perspective, where we propose a general logic-based framework for explanation generation. In particular, given a knowledge base $KB_1$ that entails a formula $\phi$ and a second knowledge base $KB_2$ that does not entail $\phi$, we seek to identify an explanation $\epsilon$ that is a subset of $KB_1$ such that the union of $KB_2$ and $\epsilon$ entails $\phi$. We define two types of explanations, model- and proof-theoretic explanations, and use cost functions to reflect preferences between explanations. Further, we present our algorithm implemented for propositional logic that compute such explanations and empirically evaluate it in random knowledge bases and a planning domain.