Inherently Interpretable Deep Reinforcement Learning Through Online Mimicking.

Although deep reinforcement learning (DRL) methods have been successfully applied in challenging tasks, their application in real-world operational settings - where transparency and accountability play important roles in automation - is challenged by methods’ limited ability to provide explanations. Among the paradigms for explainability in DRL is the interpretable box design paradigm, where interpretable models substitute inner closed constituent models of the DRL method, thus making the DRL method “inherently” interpretable. In this paper we propose a generic paradigm where interpretable DRL models are trained following an online mimicking paradigm. We exemplify this paradigm through XDQN, an explainable variation of DQN that uses an interpretable model trained online with the deep Q-values model. XDQN is challenged in a complex, real-world operational multi-agent problem pertaining to the demand-capacity balancing problem of air traffic management (ATM), where human operators need to master complexity and understand the factors driving decision making. XDQN is shown to achieve high performance, similar to that of its non-interpretable DQN counterpart, while its abilities to provide global models’ interpretations and interpretations of local decisions are demonstrated.