Catastrophe Mitigation Using DRL — LessWrong