Online Learning 3: Adversarial bandit learning with catastrophes — LessWrong