This paper addresses the computational issues involved in the solution to an infinite-horizon optimal control problem for a Markov Decision Process (MDP) with a continuous state component and a discrete control input. The optimal Markov policy for the MDP can be determined based on the fixed point solution to the Bellman equation, which can be rephrased as a constrained Linear Program (LP) with an infinite number of constraints and an infinite dimensional optimization variable (the optimal value function). To compute an (approximate) solution to the LP, an iterative randomized scheme is proposed where the optimization variable is expressed as a linear combination of basis functions in a given class: at each iteration, the resulting semi-infinite LP is solved via constraint sampling, whereas the number of basis functions is progressively increased through the iterations so as to meet some performance goal. The effectiveness of the proposed scheme is shown on a multi-room heating system example.
展开▼
