A Reinforcement Learning Approach To Adaptive Redundancy For Routing In Tactical Networks

Providing deterministic communication guarantees in dynamic and unreliable tactical networks is an ongoing challenge. Traditional routing protocols cannot adapt to the frequent topology changes inherent in battlefield scenarios, and robust flooding approaches are prohibitively expensive in terms of overhead. This paper presents a novel adaptive routing algorithm based on techniques from reinforcement learning. An online, collaborative learning algorithm gathers information about the path quality and availability to improve the packet forwarding process. Redundant routing is used when the network is highly dynamic or unknown which simultaneously increases reliability and provides ample opportunities to quickly learn the network state. This unique approach provides the benefits of learning without the drawback of a lengthy training period. The algorithm has been implemented in a Linux environment and evaluated using the CORE network emulator.