Constrained Risk-Averse Markov Decision Processes

We consider the problem of designing policies for Markov decision processes (MDPs) with dynamic coherent risk objectives and constraints. begin by formulating in a Lagrangian framework. Under assumption that constraints can be represented transition mapping, we propose an optimization-based method to synthesize Markovian lower-bound constrained risk-averse problem. demonstrate formulated optimization problems are form difference convex programs (DCPs) solved disciplined convex-concave programming (DCCP) show these results generalize linear MDPs total discounted expected costs Finally, illustrate effectiveness proposed numerical experiments on rover navigation involving conditional-value-at-risk (CVaR) entropic-value-at-risk (EVaR) measures.