Study of multi-agent systems with reinforcement learning

Collective behavior in biological systems is one of the most fascinating phenomena observed in nature. Many conspecifics form a large group together and behave col- lectively in a highly synchronized fashion. Flocks of birds, schools of fish, swarms of insects, bacterial colonies are some of the examples of such systems. Since the last few years, researchers have studied collective behavior to address challenging questions like how do animals synchronize their motion, how do they interact with each other, how much information about their surroundings do they share, and if there are any general laws that govern the collective behavior in animal groups, etc. Many models have been proposed to address these questions but most of them are still open for answers. In this thesis, we take a brief overview of models proposed from statistical physics to explain the observed collective in animals. We advocate for understanding the collective behavior of animal groups by studying the decision-making process of individual animals within the group. In the first part of this thesis, we investigate the optimal decision-making process of individuals by implementing reinforcement learning techniques. By encouraging congregation of the agents, we observe that the agents learn to form a highly polar ordered state i.e. they all move in the same direction as one unit. Such an ordered state is observed and quantified in a real flock of birds. The optimal strategy that these agents discover is equivalent to the well-known Vicsek model from statistical physics. In the second part, we address the problem of collective search in a turbulent environment using olfactory cues. The agents, far away from the odor source, are tasked with locating the odor source by sensing local cues such as the local velocity of the flow, odor plume etc. By optimally combining the private information (such as local wind, presence/absence of odors, etc.) that the agent has with public information regarding the decisions to navigate made by the other agents in the system, a group of agents complete the given search task more efficiently than as single individuals.