Goal-based multi-agent tree formation using reinforcement learning

Self - organizing natural systems are inspirational to swarm robotics. The applications include social insect systems and social behaviours. The multi - agent tree formation is one of the communication topologies of swarm robotic system. The agents are linked in a tree to communicate with each other through the hierarchy. The focus of this work is to form tree using Transfer Learning (TL) of Reinforcement Learning (RL). The goal of the approach is to maintain a certain distance with each agent. Here six agents form a tree, one among them is the leader. Each agent interacts with leader and others in the mission space till the policy of distance constraint is satisfied. The convergence of cost error is used as a paradigm to measure the knowledge transfer. Once the transfer of knowledge has taken place, the error approaches zero after two episodes and remain constant throughout the horizon. The novelty of the approach is to provide a distinct advantage when target information is unknown and only state information is known.