On Cognitive Network Channel Selection and the Impact on Transport Layer Performance

In this paper, we investigate the machine learning based strategies for dynamic channel selection in Cognitive Access Points (CogAPs) of WLANs. We employ Multi-layer Feedforward Neural Network (MFNN) models that utilize historical traffic information from network environment for learning the influence of spatio-temporal-spectral factors on the network and then predicting future traffic loads on each of the channels. Based on the future traffic loads, CogAP chooses the best channel for serving wireless clients. An important factor is the time scale of traffic prediction. We construct three kinds of traffic predictors that predict traffic at different time scales: MLP (Minute Level Prediction), MILP (Minute Interval Level Prediction), and HLP (Hourly Level Prediction) schemes and study their prediction accuracy. Experiment results show that MFNN predictors perform better than traditional autoregressive models in terms of prediction accuracy. In addition to accurate prediction, another factor that influences the design of cognitive network channel selection is the impact of channel selection strategy on the transport layer performance. We, therefore, conduct performance studies on the TCP throughput achieved on the above mentioned cognitive channel selection strategies. The MFNN predictors will also help CogAP to find and switch to the optimal channel, leading to a higher and more sustained throughput.