A Cooperative Reinforcement Learning Environment for Detecting and Penalizing Betrayal

arxiv(2022)

引用 0|浏览1
暂无评分
摘要
In this paper we present a Reinforcement Learning environment that leverages agent cooperation and communication, aimed at detection, learning and ultimately penalizing betrayal patterns that emerge in the behavior of self-interested agents. We provide a description of game rules, along with interesting cases of betrayal and trade-offs that arise. Preliminary experimental investigations illustrate a) betrayal emergence, b) deceptive agents outperforming honest baselines and b) betrayal detection based on classification of behavioral features, which surpasses probabilistic detection baselines. Finally, we propose approaches for penalizing betrayal, list directions for future work and suggest interesting extensions of the environment towards capturing and exploring increasingly complex patterns of social interactions.
更多
查看译文
关键词
cooperative reinforcement learning environment
AI 理解论文
溯源树
样例
生成溯源树,研究论文发展脉络
Chat Paper
正在生成论文摘要