An ensemble learning framework for distributed resource allocation in inteference channels: The two user case

Abstract

We focus on the problem of optimal power allocation for a two user interference channel characterized by mixed Channel State Information (CSI), which includes instantaneous information for the direct channels and statistical information for the interference channels. For this model, we introduce a general framework for optimizing the power allocation such as to maximize some generic Quality of Service (QoS) performance metric (or equivalently minimize some cost function). We model this problem as a function approximation problem where the function to be learned is the mapping between CSI and the solution to the optimization problem. We then tackle this problem borrowing ideas from ensemble learning. In particular, using generalized linear models (which are characterized by low complexity and can be implemented even at network nodes characterized by strong computational limitations), we produce different weak learners for learning to solve the considered optimization problem and based on ensemble learning theory, we combine such learners to produce stronger learners. We assess the performance of our framework by applying it on a particular resource allocation problem, and the obtained performance results indicate that the proposed approach can deliver near-optimal performance.