Reinforcement learning (RL) and especially deep reinforcement learning have emerged as powerful paradigms for learning quasi-optimal sequential decision making strategies (in applications as diverse as game playing and urban traffic control). RL uses an approach based on’ trial and error’ in which agents learn by interacting with their environment. Most RL algorithms assume that the environment is stationary, i.e., that its behaviour doesn’t change so that the learnt strategy remains optimal over time. But what happens if the environment changes (continuously) after a strategy is learnt? This project will explore how continual learning strategies might be used to update a policy as the environment changes. The project will undertake a comparative analysis of existing proposals to support continual learning and assess their applicability in a context where the system is interacting with/controlling the physical world, e.g., an automated vehicle.
School of Computer Science and Statistics