Predicting the Affordable Rate in Interference-Limited Cellular Systems Using Higher-Order Markov Models

Abstract

In cellular broadband access systems, such as 3GPP long-term evolution, the user equipment (UE) feeds back a quantized rate metric to the base station (eNodeB), in order to perform rate adaptation. The periodicity of this rate feedback is fixed so as to minimize the overhead without eroding its benefits. However, between two feedback instants n and n+δ, the actual rate that the UE can correctly decode might change due to the: Doppler shift and change in the active set of interferers. Hence, to fully exploit the benefits of adaptation, an accurate prediction of the attainable rate is required. In this context, we argue that a non-parametric approach to rate prediction is necessary. Since the selected rate is from a set of discrete values, the rate prediction problem is mapped by us onto a discrete sequence prediction problem, and we construct higher order Markov models for the discrete sequences using source encoding algorithms. We then propose two distinct rate-prediction algorithms. One of them is the adaptive maximum a posteriori estimator, while the other is the adaptive Bayesian risk-based estimator. Both of these algorithms simultaneously estimate the best Markov model for each UE and then perform prediction based on the estimated model.

Keywords

• Statistical learning
• adaptive algorithms
• prediction algorithms
• feedback communications
• adaptive estimation
• time series analysis