Evaluation of HetNet Offloading with a GAN-Driven Wireless Testbed

We propose a wireless traffic workload generator for use in simulations and testbeds to evaluate machine learning based system optimizations without real users. Our generator produces uplink and downlink load traces for any number of users and any duration by utilizing a user-aware trained generative adversarial network (GAN) while preserving the key statistical features and privacy of the original training data. We demonstrate the use of the generator in a testbed evaluation of a neural network model for HetNet device offloading. Our experiment testbed replays these generated traces across a set of offloading strategies in order to determine the optimum set of users to move off a Wi-Fi network to a LiFi network in an Enterprise office setting. Our results demonstrate how a neural network model significantly outperforms simpler strategies in terms of network KPI, such as air utilization and collision probability.