Abstract
Model-free reinforcement learning has seen tremendous advances in the last few years, however practical applications of pure reinforcement learning are still limited by sample inefficiency and the difficulty of giving robustness and stability guarantees of the proposed agents. Given access to an expert policy, one can increase sample efficiency by in addition to learning from data, and also learn from the experts actions for safer learning.
In this paper we pose the question whether expert learning can be accelerated and stabilized if given access to a family of experts which are designed according to optimal control principles, and more specifically, linear quadratic regulators. In particular we consider the nominal model of a system as part of the action space of a reinforcement learning agent. Further, using the nominal controller, we design customized reward functions for training a reinforcement learning agent, and perform ablation studies on a set of simple benchmark problems.