Techniques for optimizing worst-case performance — LessWrong