Explainable Deep Reinforcement Learning through Introspective Explanations

Nowadays modern life is influenced in many ways by artificial intelligence systems. This could be a decision of an artificial intelligence, an autonomous vehicle or functions of consumer devices. In many domains artificial intelligence can act faster, more precise or better than a human could. At the same time, most decision processes of artificial intelligence systems are opaque black box systems. Making these decision processes understandable to humans could increase the reliability, maintainability and trustworthiness of such systems. (World Model based) Deep Reinforcement Learning Agents are often too complex to be explained because decisions are influenced by previous states and actions. To address this complexity, a new multistep explanation will be introduced. The proposed approach uses learned world models and policies to imagine the future states and predicting actions for them. The imagination is used as explanation to understand the plan of the agent and to know and validate the targeted goal. A user study design was created to evaluate the effectiveness of this new form of explanation.