Integrating teletraffic theory with neural networks for quality-of-service evaluation in mobile networks

Abstract

In mobile cellular design, one important quality-of-service metric is the blocking probability. Using computer simulation for studying blocking probability is quite time-consuming, whereas existing teletraffic-based methods such as the Information Exchange Surrogate Approximation (IESA) only give a rough estimate of blocking probability. Another common approach, direct blocking probability evaluation using neural networks (NN), performs poorly when extrapolating to network conditions outside of the training set. This paper addresses the shortcomings of existing teletraffic and NN-based approaches by combining both approaches, creating what we call IESA-NN. In IESA-NN, an NN is used to estimate a tuning parameter, which is in turn used to estimate the blocking probability via a modified IESA approach. In other words, the teletraffic approach IESA still forms the core of IESA-NN, with NN techniques used to improve the accuracy of the approach via the tuning parameter. Simulation results show that IESA-NN performs better than previous approaches based on NN or teletraffic theory alone. In particular, even when the NN cannot produce a good value for the tuning parameter, for example when extrapolating to network conditions not experienced in the training set, the final IESA-NN estimate is generally still accurate as the estimate is primarily determined by the underlying teletraffic theory, with the NN determining the tuning parameter playing a supplementary role. The combination of the IESA framework with NN in a secondary role makes IESA-NN quite robust.