Learning-based task allocation in decentralized multirobot systems

This paper presents an efficient decentralized multirobot system through the use of reinforcement learning coupled with heuristics to accelerate the learning process. Our heuristic functions are not reinforcement signals, but are biases to the learning space that help accelerate the exploration process and have no effect on the robot's knowledge. Two kinds of heuristics are used: local and global. The local heuristic is derived from the local knowledge of the robot, whereas the global heuristic is derived from the data obtained from other robots. Both heuristics can help accelerating the learning process.