b) Decoupling the action elimination (by the AEN) from the action selection (by the DRL agent) reduces the risk of the model exploring only a small subset of actions.