
S2AC learns expressive stochastic policies modeled as Stein Variational Gradient Descent (SVGD) samplers from Energy-Based Models (EBMs) over Q-values, enabling it to maximize both the expected future reward and the expected future entropy in a computationally efficient manner.


coremsg

stein-soft-actor-critic-an-energy-based-reinforcement-learning-algorithm-with-expressive-stochastic-policies


Stein Soft Actor-Critic: An Energy-Based Reinforcement Learning Algorithm with Expressive Stochastic Policies
