Predicting Bit Error Rate from Meta Information using Random Forests and Partial Data Augmentation

With the increasing power of machine learning-based reasoning, the use of meta information (e.g., digital signal modulation parameters, channel conditions, etc.) to predict the performance of various signal processing techniques has become feasible. One such problem of practical interest is choosing a proper interference mitigation method based on the meta information of the received signal. Since heuristic table-based methods suffer from limited prediction capability for unseen cases, we propose a recommendation system based on the use of Random Forests (RF). Specifically, RF are used to predict the Bit-Error-Rate (BER) of all mitigation approaches so as to determine the approach with the best performance. In this work we show that RF can predict BER with high accuracy, and the Importance Factor (IF) demonstrates the input attributes which matter most. Furthermore, the IF can help to design a data augmentation of partial features. These BER prediction results can also benefit other functions such as adaptive modulation, channel sensing, beam selection, etc.