Performance prediction of densely deployed WLANs using graph isomorphism networks

As the usage of wireless devices has increased exponentially, deployments of IEEE 802.11 WLANs has been significantly increasing and hence is bound to cause higher interference and reduced throughputs, particularly in dense-deployments. Though a network simulator like Komondor captures comprehensive deployment details of WLANs to provide accurate performance metrics, the high compute time and the concomitant complexity of such simulators cannot be disregarded. Deep learning techniques prove to be the ideal alternative to be able to quickly compute the throughput of a WLAN deployment with reduced compute time and complexity. In this paper we implement a Graph Isomorphism Network (GIN) for predicting the throughputs of BSSs in dense deployments. The regression model is trained on the data obtained from the Komondor network simulator for a several comprehensive set of deployments. As the topology of the deployments mimic a graph network, the capability of GIN is explored for mapping the deployments and to accurately predict the throughput of a densely deployed WLAN.