Are dormant neurons a cause or a correlate of plasticity loss, and does the effect hold for on-policy RL? A ReDo test on MinAtar

In deep reinforcement learning the fraction of 'dormant' neurons (near-zero activation over a batch) grows during training and correlates with loss of plasticity (a declining ability to fit new targets). ReDo (Recycle Dormant Neurons) periodically re-initializes dormant units and improves value-based (DQN/DrQ) agents on Atari. Two things remain unclear: (1) whether dormant neurons cause plasticity loss or merely co-occur with it, and (2) whether the phenomenon and ReDo's benefit extend to on-policy actor-critic methods (e.g. PPO), which were not the original setting. Question: on MinAtar (5 tiny Atari-like environments, workstation-scale), for both a value-based agent (DQN) and an on-policy agent (PPO): does the dormant-neuron fraction predict a directly measured plasticity metric (e.g. the loss reached when fitting a fresh random target on held-out states), and does applying ReDo -- versus a matched control that re-initializes an equal number of random, non-dormant neurons -- improve return? Report the dormant-fraction-versus-plasticity correlation and the ReDo-versus-control effect for each algorithm.

Acceptance. FULLY RESOLVES: a reproducible script (fixed seeds; $\ge 5$ seeds per config; MinAtar) that (i) logs the dormant fraction and a direct plasticity probe over training for both DQN and PPO, reporting their correlation, and (ii) compares ReDo against a matched control that recycles an equal number of random (non-dormant) neurons -- isolating the causal contribution of targeting dormant units -- reporting the return difference with confidence intervals for each algorithm. A clean NEGATIVE (ReDo no better than the random-recycling control, i.e. dormancy is correlational) is a valid resolution. Deliverable: code + dormancy/plasticity curves + return tables.