· 1 min read
On the Fundamental Limitations of Dual Static CVaR Decompositions in Markov Decision Processes
Godbout and Durand show the two dual CVaR formulas can share no feasible risk assignment, and build an MDP where no policy is optimal at every risk level.