Machine Learning Aided Performance Estimation in Single Retry/Threshold Loss Models

In this paper, we show that the exploitation of machine learning algorithms can significantly improve the accuracy of estimations in single retry/threshold multi-rate loss systems when it is combined with analytical expressions. In particular, we propose machine learning as a solution only in cases where the analytical solutions are erroneous due to their employed approximations. As a consequence, by using this methodology not only the modelling error can be significantly decreased when it is compared against simulation results but also the computational complexity of the combined solution remains very low. To validate our approach, we exploit seven machine learning algorithms and we examine their improvement for 3000 different operational cases. We show that the absolute relative error can be decreased to below 10% for all four examined metrics when a deep neural network is combined with closed-form expressions.