Curiosity Killed or Incapacitated the Cat and the Asymptotically Optimal Agent

Reinforcement learners are agents that learn to pick actions lead high reward. Ideally, the value of a reinforcement learner's policy approaches optimality-where optimal informed is one which maximizes Unfortunately, we show if an agent guaranteed be “asymptotically optimal” in any (stochastically computable) environment, then subject assumption about true this will either “destroyed” or “incapacitated” with probability 1. Much work learning uses ergodicity avoid problem. Often, doing theoretical research under simplifying assumptions prepares us provide practical solutions even absence those assumptions, but may have led entirely astray preparing safe and effective exploration strategies for dangerous environments. Rather than assuming away problem, present agent, Mentee, modest guarantee approaching performance mentor, instead reckless exploration. Critically, Mentee's depends on expected information gain from exploring. In simple non-ergodic environment weak find Mentee outperforms existing asymptotically its mentor.