A Multi-agent Reinforcement Learning Routing Protocol in Mobile Robot Network

Robots are now essential in unreachable, repeated, and dangerous real-world applications where they take place of human beings. One of the important capabilities of a multi-robot system is that it should be able to form an autonomous robot network to transmit information. However, due to the limited communication capability of a single robot, the highly variable environment where robots work, and the mobility of the robots, it is difficult for them to exchange information with each other in need. In this article, we propose a novel robot network routing protocol based on multi-agent reinforcement learning called MAQR. The robot nodes can deliver packets cooperatively. The mobility factor, buffer status, and packet delay of neighbor nodes are taken into consideration. We design a reliability model for a robot agent to make reliable routing decisions. We also design an adaptive exploration-exploitation method to balance the convergence speed, solution space as well as network fluctuation. The routing algorithm has been implemented and evaluated in simulation. Results show that MAQR can provide less packet delay, less average queue length and higher delivery ratio than other Q-learning based routing protocol.