Curriculum Reinforcement Learning for Cohesive Team in Mobile Ad Hoc Networks

초록

In emergency scenarios, such as disaster or military situations, ad hoc networks should be deployed as no central coordination is available. In this letter, we propose a distributed solution for building mobile ad hoc networks, where the mobile nodes determine their positions as a team autonomously based on reinforcement learning. We propose a special design of a decentralized partially observable Markov decision process to build a cohesive team of mobile nodes in a distributed manner. Each mobile node in the team learns an individual policy that determines movement under partial observation, with the common goal of maximizing network throughput. In the learning process, each node indirectly negotiates the role in the team while explicitly considering the locations of other neighboring nodes and network throughput. To improve learning efficiency, we design a curriculum that encourages nodes to disperse initially but reside in specific regions eventually. Such a curriculum enables each node to be placed in its best location, thereby expediting the collective convergence of all nodes as a cohesive team. Simulation results confirm that the proposed solution can successfully build a cohesive team that maintains high network throughput with low power consumption. © 1997-2012 IEEE.

키워드