Multiagent Reinforcement Learning by exchanging observations and commitments (FWOTM1008)

Multiagent systems represent a powerful tool for modeling distributed settings that require robust, scalable, and decentralised control solutions. The complexity of those systems makes them in general hard to engineer. We consider here the multiagent Reinforcement Learning (RL) as an interesting paradigm for such systems. Despite this field receiving growing attention during the last decade, there are still important unexplored research directions. Recent results exposed the fragility of state-of-the-art algorithms when introducing perturbations during the execution of the learned behavior. Aware of this flaw, we aim to develop more robust algorithms. As no environments have been designed with this perspective in mind, we will first produce a suite of environments to benchmark the robustness of multiagent RL algorithms. For agents to learn more robust policies, they need to anticipate perturbations that might happen when performing the learned policy. We will develop novel algorithms that are able to identify crucial observations and share them between agents, and that allow anticipation to improve their robustness in both fully-cooperative and mixed settings. We will extend our solutions such that agents are able to identify that they are facing unseen perturbations in order to appropriately react on this. Furthermore, we will explore ideas from model-based RL, to allow agents to commit about future states, which will enable the multiagent system to learn better solutions.

Keywords:Multiagent Reinforcement Learning, Distributed Deep Learning

Disciplines:Adaptive agents and intelligent robotics, Machine learning and decision making