Reactivity and safe learning in multi-agent systems

Multi-agent reinforcement learning (MRL) is a growing area of research. What makes it particularly challenging is that multiple learners render each other's environments non-stationary. In addition to adapting their behaviors to other learning agents, online learners must also provide assurances about their online performance in order to promote user trust of adaptive agent systems deployed in real world applications. In this article, instead of developing new algorithms with such assurances, we study the question of safety in online performance of some existing MRL algorithms. We identify the key notion of reactivity of a learner by analyzing how an algorithm (PHC-Exploiter), designed to exploit some simpler opponents, can itself be exploited by them. We quantify and analyze this concept of reactivity in the context of these algorithms to explain their experimental behaviors. We argue that no learner can be designed that can deliberately avoid exploitation. We also show that any attempt to optimize reactivity must take into account a tradeoff with sensitivity to noise, and devise an adaptive method (based on environmental feedback) designed to maximize the learner's safety and minimize its sensitivity to noise.