Efficient and Fair Decentralized Task Allocation Algorithms for Autonomous Vehicles: A Machine Learning Based Approach

The objective of this project is to improve the efficiency of the multi-agent decentralized mission coordination with an inter-agent communication infrastructure. In phase 1 of this project, we explored the enhancement of the Consensus Based Bundle Algorithm (CBBA) for distributed task allocation with budget constraints. The limitation of the CBBA technique is that the environment must be known a priori to all agents and tasks must be clearly defined with known costs and rewards. This technique is obviously not suitable for cooperative missions in an unknown environment where agents must explore and improvise their actions together. In phase 2 of this project, we study cooperation techniques for missions in unknown environment where agents have only partial observations. The study uses multi-agent predator and prey game as a platform. The goal is for the agents to jointly locate and capture the prey. The agents have no prior knowledge of the environment or the preys escape algorithm. They communicate with each other to obtain environment information beyond their own local observations. Based on their local understanding of the environment, the agents choose their own action, which includes where to move and whether to communicate with other agents, to maximize the team rewards. Reinforcement learning is applied to optimize the agents policy such that the game is completed with the fewest steps.