Reliability Enhancement of 3G Radio Network Prediction by a Conditional Distribution Discrimination Tree

This paper presents a constructive learning system that enhances the reliability and precision of radio network predictions. This task is achieved by finding a correspondence between the probability density distributions of simulated predictions and real measurement data collected from the radio network. Once this correspondence is found, it is possible to arrive at more realistic prediction values from simulation results. After carrying out non-parametric estimations of the probability distributions of the simulations, feature vectors are computed from these estimations, followed by a supervised learning that finds a mapping between the feature vectors issued from the simulations and the estimations of conditional probability distributions of the measurements. The proposed method is evaluated on a 3G radio network using indicators such as UpLink (UL) and DownLink (DL) base station loads. Results show that the proposed scheme is able to yield distributions that are much closer to measurements than simulations. With such a technique, it is possible to predict with enhanced accuracy new configurations and conditions for which we don´t have observations.