ConceptEntropy-Regularized RL
Entropy-Regularized Reinforcement Learning
A reinforcement learning setting in which an entropy regularization term is added to the reward function. Its optimal policy takes the form of a Gibbs distribution, and the coefficient of the entropy term is the reciprocal of the temperature. It serves as the basis for the tilted sampling framework to unify multiple settings.
Timeline (last 90 days)
Oct 16
In entropy-regularized reinforcement learning, the optimal policy is exactly of the Gibbs distribution form, and the coefficient of the entropy regularization term and the temperature are reciprocals of each other, making maximizing the entropy-regularized reward mathematically equivalent to matching the Gibbs distribution.
Unverified50%