CMDP

An extension of the Markov decision process framework that requires maximizing return while satisfying threshold constraints on cumulative cost. It is the classical formalization of safe reinforcement learning.