Online Learning and Optimization of Markov Jump Affine Models

Abstract: The problem of online learning and optimization of unknown Markov jump affine models is considered. An online learning policy, referred to as Markovian simultaneous perturbations stochastic approximation (MSPSA), is proposed for two different optimization objectives: (i) the quadratic cost minimization of the regulation problem and (ii) the revenue (profit) maximization problem. It is shown that the regret of MSPSA grows at the order of the square root of the learning horizon. Furthermore, by the use of van Trees inequality, it is shown that the regret of any policy grows no slower than that of MSPSA, making MSPSA an order optimal learning policy. In addition, it is also shown that the MSPSA policy converges to the optimal control input almost surely as well as in the mean square sense. Simulation results are presented to illustrate the regret growth rate of MSPSA and to show that MSPSA can offer significant gain over the greedy certainty equivalent approach.