An online learning framework for link adaptation in wireless networks

Current and future wireless networks require the selection of a plurality of parameters at different layers of the communication system to optimize network throughput while satisfying certain reliability constraints. In prior work, mathematical input/output models along with system performance expressions have been used to perform the parameter selection. In practice, however, impairments such as interference and analog circuit nonlinearities are difficult to model in a simple and tractable framework. Moreover, these impairments are in flux due to environmental factors. This paper summarizes an online machine learning approach to parameter selection through the real-time capturing of performance related data. Online learning is advantageous, not only because changes in the system model can be captured in the data observations, but also for its ability to learn system operation details not provided by current system models. A modified version of k-nearest neighbor is developed to enable both the real-time capture of training data and the performance criterion for physical layer adaptation.