Learning by doing and the value of optimal experimentation

Research on learning-by-doing has typically been restricted to cases where estimation and control can be treated separately. Recent work has provided convergence results for more general learning problems where experimentation is an important aspect of optimal control. However the associated optimal policy cannot be derived analytically because Bayesian learning introduces a nonlinearity in the dynamic programming problem. This paper characterizes the optimal policy numerically and shows that it incorporates a substantial degree of experimentation. Dynamic simulations indicate that optimal experimentation dramatically improves the speed of learning, while separating control and estimation frequently induces a long-lasting bias in the control and target variables.