Tag
This paper analyzes over 33,000 experiments on actor-critic algorithm design components, finding that common defaults like Gaussian action distributions are unreliable while bounded distributions with adaptive update schedules are robust, offering practical guidance for practitioners.