Reinforcement learning based routing in wireless mesh networks

This paper addresses the problem of efficient routing in backbone wireless mesh networks (WMNs) where each mesh router (MR) is equipped with multiple radio interfaces and a subset of nodes serve as gateways to the Internet. Most routing schemes have been designed to reduce routing costs by optimizing one metric, e.g., hop count and interference ratio. However, when considering these metrics together, the complexity of the routing problem increases drastically. Thus, an efficient and adaptive routing scheme that takes into account several metrics simultaneously and considers traffic congestion around the gateways is needed. In this paper, we propose an adaptive scheme for routing traffic in WMNs, called RLBDR (Reinforcement Learning-based Distributed Routing), that (1) considers the critical areas around the gateways where mesh routers are much more likely to become congested and (2) adaptively learns an optimal routing policy taking into account multiple metrics, such as loss ratio, interference ratio, load at the gateways and end-to end delay. Simulation results show that RLBDR can significantly improve the overall network performance compared to schemes using either interference and channel switching (MIC), Best Path to Best Gateway (BP2BG), Expected Transmission count (ETX), nearest gateway (i.e., shortest path to gateway) or load at gateways as a metric for path selection.