Dual control for approximate Bayesian reinforcement learning (Q2834442)

Bayesian reinforcement learning, or dual control, provides a principled solution to the exploration-exploitation trade-off for learning in interactive settings. This paper extends an old Bayesian reinforcement learning in control theory in the context of modern regression methods with ideas from contempary machine learning, including approximate Gaussian process regression and multi-layer networks. Experimental results and simple examples are also given to demonstrate the effectiveness of the proposed dual control framework.