A dynamical neural network approach for distributionally robust chance constrained Markov decision process
In this paper, we study the distributionally robust joint chance constrained Markov decision process. {Utilizing the logarithmic transformation technique,} we derive its deterministic reformulation with bi-convex terms under the moment-based uncertainty set. To cope with the non-convexity and improve the robustness of the solution, we propose a dynamical neural network approach to solve the reformulated optimization problem. Numerical results on a machine replacement problem demonstrate the efficiency of the proposed dynamical neural network approach when compared with the sequential convex approximation approach.