Deep reinforcement learning for plan execution

There are many different methods for the deliberative control of autonomous systems in stochastic environments, each with different strengths and limitations. Reinforcement Learning can provide robust performance in unpredictable environments, but its decisions are often not predictable. In contrast, Automated Planning can provide explicable and transparent behaviour but its performance drops when the environment is uncertain. In this paper we discuss an approach to plan execution through reinforcement learning by training an agent to follow predetermined plans. The implementation of the approach leads to the complex task of defining evaluation metrics that describe the desired behaviour. We describe the implementation of this approach as a set of agents, which differ in their reward function, and were trained and evaluated in three scenarios in which plan execution can deviate and be recovered.