Certified data-driven inverse reinforcement learning of Markov jump systems

This paper devises a data-driven off-policy inverse reinforcement learning algorithm for discrete-time linear Markov jump systems. Leveraging observed behaviors from multi-mode Markov jump systems optimized with respect to unknown mode-wise cost functions, our algorithm reconstructs each mode’s associated cost function and control policies. Our approach certifies mean-square asymptotic stability, unbiased convergence, and nonuniqueness. Notably, the proposed algorithm operates solely on demonstrated behavioral data, eliminating the need for system models, prior knowledge of transition probabilities, or initial control gains. The practical efficacy of our approach is validated.