Abstract
Recent advances in reinforcement learning (RL) have enabled data-driven controller design for autonomous systems. Applying RL typically requires encoding the objective using a reward function and then training a neural network controller to maximize the expected reward. However, manually encoding complex objectives as low-level, state-based reward functions can be tedious and error-prone. Furthermore, the choice of encoding can impact training time by orders of magnitude. This motivates research on using specification languages, such as Linear Temporal Logic (LTL), that make it easier to specify complex RL tasks in an intuitive yet precise manner. This article gives a tutorial-style introduction to representative results on algorithmic foundations of specification-guided RL.
