ETD

• continual learning
• control tasks
• evolution strategies
• lifelong learning

This work investigates the behaviour of Evolution Strategies (ES) in continual learning settings, analysing their capacity to retain, transfer, and adapt knowledge across sequential reinforcement learning tasks. Through systematic experiments on MuJoCo control environments, the study quantifies forgetting and transfer under different architectural and replay configurations. Results show that replay mechanisms effectively stabilise learning and substantially mitigate catastrophic forgetting, performing reliably up to the limits of the studied problems. The analysis highlights that most interference arises in task-specific output layers and that modular architectures further enhance stability. Overall, this thesis provides the first detailed empirical characterisation of continual learning dynamics in Evolution Strategies, establishing a foundation for future research on adaptive and scalable gradient-free continual optimisation.