Entropy Objectives in Markov Decision Processes

The paper presents a formal approach to synthesizing control policies in Markov Decision Processes (MDPs) that maintain an entropy-based objective, highlighting the complexity of even relaxed versions of this problem. It introduces a sound and conditionally complete method for verifying and synthesizing strategies, leveraging convex duality and invariant synthesis to tackle the non-linear nature of entropy objectives. This work is significant for practitioners as it provides a framework for implementing entropy constraints in stochastic systems, potentially enhancing the robustness and performance of AI decision-making processes.