Collision avoidance by mitigating uncertain packet loss in multi-hop wireless IoT networks

Multi-hop wireless relaying is an effective solution to provide connectivity to IoT devices in places that are difficult to reach. Spatial reuse for higher spectral efficiency by allowing simultaneous transmissions, however, causes self-interference unless transmissions are carefully coordinated. To solve this issue, recently, ML(Machine Learning)-based transmission scheduling has been explored in many literatures. Existing ML- based schemes, however, have limitation in that they do not account for the control overhead associated with schedule deployment and network state collection. In this paper, we propose a DRL (Deep Reinforcement Learning)-based TDMA scheduling scheme that aims to optimize network throughput and minimize energy consumption while avoiding collisions. More specifically, we use a Sequence-to-Sequence (S2S) neural network to compose the DRL policy. One of the key novelties of our scheme is that the schedule deployment is conducted sparsely to reduce the control overhead. This causes uncertainties due to the random packet losses, and we mitigate the uncertainties via a technique called redundant scheduling. Simulation results demonstrate that the proposed scheme is scalable and converges quickly, and it outperforms existing schemes under various network conditions.