The Soft Actor-Critic (SAC) approach to discovering policies with real-valued actions under reinforcement learning tasks employs entropy regularization to balance the exploration—exploitation trade-off. The amount of regularization is controlled by a hyper-parameter, \(\alpha \) . Contemporary approaches for annealing \(\alpha \) concentrate on linking adaption to previously encountered actions and rewards. We introduce a small neural network to specifically tune \(\alpha \) . This enables us to also explicitly anneal \(\alpha \) as a function of task state. All other aspects of the SAC framework remain unchanged, making the approach easy to deploy with current SAC code bases. Benchmarking on four MuJoCo locomotion tasks demonstrates unique dependencies between the annealed \(\alpha \) trajectory, type of task and state while matching/bettering reward cumulation from other contemporary methods (DDPG,PPO, TD3). Moreover, hyper-parameter optimization for \(\alpha \) is avoided.