Cynnovative, LLC — Department of Defense STTR Phase I: MDA22-T004

Description

Cynnovative proposes Explainable Deep Reinforcement Learning with Symbolically Guided Transitions (X DRLSGT) to improve the transparency and, thus, the explainability of deep reinforcement learning (DRL) algorithms. The inability to understand the reasoning behind an Artificial Intelligence’s (AI) decision is a major limiting factor that prevents AI-enabled physical systems from being deployed alongside humans. This is especially true for the warfighter and analysts who must regularly manage high volumes of information at any given time and who could benefit significantly from working with an AI, which could process large amounts of information quickly and expose only the necessary elements to the operator. Our proposed approach will improve explainability in DRL by extracting meaningful and transparent information from the model that the agent uses to reason about the world. We will accomplish this by leveraging a model-based reinforcement learning algorithm that learns to represent the world in a way that can be used to ultimately derive meaningful information from the mind of the agent. This will provide insight into how the agent sees the world and how it expects the world to change. Approved for Public Release | 22-MDA-11339 (13 Dec 22)