Gaussian processes for informative exploration in reinforcement learning

This paper presents the iGP-SARSA(λ) algorithm for temporal difference reinforcement learning (RL) with non-myopic information gain considerations. The proposed algorithm uses a Gaussian process (GP) model to approximate the state-action value function, Q, and incorporates the variance measure from the GP into the calculation of the discounted information gain value for all future state-actions rolled out from the current state-action. The algorithm was compared against a standard SARSA(λ) algorithm on two simulated examples: a battery charge/discharge problem, and a soaring glider problem. Results show that incorporating the information gain value into the action selection encouraged exploration early on, allowing the iGP-SARSA(λ) algorithm to converge to a more profitable reward cycle, while the e-greedy exploration strategy in the SARSA(λ) algorithm failed to search beyond the local optimal solution.