Dynamic programming for ergodic control with partial observations. (Q2574544): Difference between revisions

The paper derives a dynamic programming principle for optimal control of a partially observed Markov process taking values in a Euclidean space. The minimized functional is that of average (ergodic) costs over infinite horizon. The control space is compact. The problem is addressed by approximating the original ergodic cost functional by a family of discounted cost functionals with discount factors converging to unity. The dynamic programming principle inequalities are first derived in discrete time and the result is then carried over to partially observed Markov semimartingales in continuous time. The construction of optimal controls proceeds in the following steps: 1. restating the problem by means of a separation principle which makes the control process adapted to the process of observations, 2. changing the probability measure in order to eliminate variability in the marginal distribution of the observation process, 3. introducing a stability assumption for the state process in a Lyapunov function form, 4. embedding the state process into another one with a ``doubled'' range of values, for which an accessible atom exists. The argument draws on earlier results of the same author concerning optimal ergodic control of partially observed finite Markov chains.