Episodic Memory: Foundation of Explainable Autonomy

Abstract

In recent years, the inner workings of many intelligent agents have become opaque to users who wish to manage or collaborate with them. This lack of transparency makes it difficult for human users to understand and predict the behavior of such agents. We argue that computational agents that store the plans they construct can behave in a predictable and transparent manner by remembering the plans used for achieving goals and explaining their contents to users. To investigate this issue, we present a psychologically inspired computational theory of episodic memory that explains how intelligent agents can use their personal experience to make known their internal decision-making process. We augment this theory with an implementation and show how systems with episodic memory capabilities can explain what happened in their personal past. We demonstrate this system's ability to answer questions in two Minecraft scenarios. Our preliminary findings suggest that episodic memory capabilities in computational agents plays an important role in producing explanations regarding an agent's cognitive and behavioral abilities. With continued research, we believe our approach can facilitate the harmonious integration of robots with their human counterparts, creating an environment where humans and artificial agents better understand each other.