Cooperative routing for parallel data transmissions in Internet of UAVs: a deep reinforcement learning approach

Abstract

The Internet of unmanned aerial vehicles (UAVs) is a flying ad hoc network (FANET) of multiple UAVs connected via wireless links, where the UAV nodes share information and cooperate with each other through real-time communication. To improve the system capacity, the data packets of different links transmit in parallel where some of the UAVs work as relaying nodes to forward data. Due to the interference of wireless channels and the dynamic topology of the network, parallel data routing is challenging in the Internet of UAVs. Traditional routing protocols may fail for three reasons: the wireless interference decreases the channel capacity; the dynamic topology makes it harder to keep the routing table; parallel data routing may cause path collision. As a result, we propose a cooperative routing approach for parallel data transmissions in the Internet of UAVs with a deep reinforcement learning algorithm. The scenario can be modeled as a multi-agent path finding (MAPF) problem where each data packet can be treated as an agent. Specifically, each agent observes and evaluates its adjacent nodes’ features instead of keeping routing tables and selects one as the next hop until arriving at the destination. In order to encourage cooperation between the agents, the reward of their actions is set as the increment of system throughput rather than the rate of their own paths. With centralized training and decentralized execution (CTDE), the agents learn to cooperatively relay data with the deep q-learning (DQN) algorithm. Compared to traditional routing protocols in wireless ad hoc networks such as the ad hoc on-demand distance vector routing (AODV), our proposed algorithm can significantly improve system throughput and shorten the number of relay hops.