Multi-Agent Deep Learning for Multi-Channel Access in Slotted Wireless Networks

As the number of devices connected to the internet and the amount of data they generate increases, the wireless spectrum is becoming an essential and scarce resource. Most connected devices use wireless technologies that use the industrial, scientific, and medical (ISM) radio bands. As a result, different technologies are interfering with each other. Today's existing collision avoidance techniques either apply a random back-off when a signal collision is detected or assume that knowledge about other nodes' spectrum occupation is known. These approaches are competent approaches to optimise inter-network spectrum usage, but fail to optimise overall channel capacity and throughput of all neighbouring wireless networks. In this paper, we present a Deep Neural Network (DNN) approach that can predict spectrum occupation of unknown neighbouring networks in the near future by using online supervised learning in a multi-agent setting. This prediction can be employed by existing network schedulers to avoid collisions with surrounding networks or other electromagnetic sources. The DNN is trained in an online way, as the problem is a partially observable stochastic game with continuous action space. Our findings show a reduction in the number of collisions between the own network and neighbouring networks of 30%, and an increase in overall throughput of 10% in a medium-sized network with an unknown set of neighbouring networks.