Lumped Markovian Estimation for Wi-Fi Channel Utilization Prediction

We present a model to predict the short-term utilization of an IEEE 802.11 channel. We approximate the time-varying utilization process via a Markovian state transition model and subsequently create a lumped representation of the transition matrix. Each lumped state can then be treated as a class. The lumped matrix provides a simpler to understand description of the channel utilization behavior and naturally includes the persistence in one lumped state which resembles the characteristic behavior of naive predictors (where predicted state equals the current state). We demonstrate that treating the lumped states as classes allows good prediction models to be built using Logistic Regression and Neural Network models. Our results are based on IEEE 802.11 wireless utilization data collected as reported in the channel utilization (CU) field of the QBSS Load Element in Beacon frames. The presented approach can be implemented as an edge computing task, whereby edge nodes calculate the lumped states and train models, informing nearby client devices of the model parameters, allowing them to produce predictions on their own.