Tag
This paper analyzes the Option-Critic framework in reinforcement learning, explaining that the learned termination rule is redundant, introducing 'policy necrosis' where options fail to explore, and showing that extra options improve performance by reducing the probability of joint failure.