Towards an Unified Structure for Reinforcement Learning: an Optimization Approach

Abstract: Both the optimal value function and the optimal policy can be used to model an optimal controller based on the duality established by the Bellman equation. Even with this duality, no parametric model has been able to output both policy and value function with a common parameter set. In this paper, a unified structure is proposed with a parametric optimization problem. The policy and the value function modelled by this structure share all parameters, which enables seamless switching among reinforcement learning algorithms while continuing to learn. The Q-learning and policy gradient based on the proposed structure is detailed. An actor-critic algorithm based on this structure, whose actor and critic are both modelled by the same parameters, is validated by both linear and nonlinear control.