Topology-Transparent Scheduling Based on Reinforcement Learning in Self-Organized Wireless Networks

Topology-transparent scheduling policies do not require the maintenance of accurate network topology information and therefore are suitable for highly dynamic scenarios in self-organized wireless networks. However, in topology-transparent scheduling, it is a very challenging problem to make individual nodes efficiently select their transmission slots in a distributed manner. It is desirable for individual nodes, through time slot selection, to avoid collision on the one hand and utilize as many time slots as possible (i.e., minimize the number of redundant slots) on the other. In this paper, learning-based approaches are employed to solve the time slot scheduling problem. Specifically, the proposed method uses a temporal difference learning approach to address the collision issue and use a stochastic gradient descent approach to reduce the number of redundant slots. Unlike previous works, this learning approach is trained through self-play reinforcement learning without incurring communication overhead for the exchange of reservation information, thereby improving the network throughput. Extensive simulation results validate that our proposal can achieve better efficiency than the existing approaches.