Generalizing POWER to multi-agent games — LessWrong