Learning and control in a changing economic environment

In this paper we investigate optimal Bayesian learning and control with lagged dependent variables and time-varying unknown parameters. We assess both the performance of alternative decision rules, as in the computationally-oriented dual control literature, as well as the dynamics of learning and convergence of beliefs, as in the more theoretically oriented Bayesian learning literature. Our numerical results indicate that the optimal decision rule involves a noticeable degree of experimentation for moderate to large levels of uncertainty. In most situations though, the optimal rule will remain less activist than a certainty-equivalent rule and induce gradualism, Exceptions occur when the process to be controlled is near the deterministic steady state. In this situation, we find that the decision-maker will repeatedly undertake costly experiments. The extent of optimal experimentation is lower if the unknown parameter is perceived to vary over time. In contrast to the fixed parameter case, however, parameter uncertainty is continually renewed and the incentive to experiment never ceases. (C) 2002 Elsevier Science B.V. All rights reserved.