Plasticity loss is a real problem in RL, but its severity depends on task difficulty and noise levels—noise can help or hurt learning depending on the environment, and early training is when plasticity loss matters most.
This paper investigates how neural networks lose their ability to learn (plasticity loss) during deep reinforcement learning in noisy environments. The authors train SAC agents on control tasks with varying noise levels and find that plasticity-preserving methods improve robustness on harder tasks, while noise can either hurt learning or act as a regularizer depending on task difficulty.