Covert Planning

Covert planning focuses on designing agents that accomplish tasks while minimizing information leakage to observers, thus avoiding detection. Current research emphasizes using reinforcement learning, particularly Markov Decision Processes (MDPs) and algorithms like Conservative Q-Learning and policy gradient methods, to optimize agent actions for both task completion and covertness, often incorporating stochasticity and imperfect observer models. This field is significant for applications like secure distributed optimization (e.g., federated learning) and autonomous robot navigation in adversarial environments, offering solutions for enhancing privacy and security in various domains.