Towards Data-Driven Simulation of End-to-End Network Performance Indicators

Abstract

Novel vehicular communication methods are mostly analyzed simulatively or analytically as real world performance tests are highly time-consuming and cost-intense. Moreover, the high number of uncontrollable effects makes it practically impossible to reevaluate different approaches under the exact same conditions. However, as these methods massively simplify the effects of the radio environment and various cross-layer interdependencies, the results of end-to-end indicators (e.g., the resulting data rate) often differ significantly from real world measurements. In this paper, we present a data-driven approach that exploits a combination of multiple machine learning methods for modeling the end-to-end behavior of network performance indicators within vehicular networks. The proposed approach can be exploited for fast and close to reality evaluation and optimization of new methods in a controllable environment as it implicitly considers cross-layer dependencies between measurable features. Within an example case study for opportunistic vehicular data transfer, the proposed approach is validated against real world measurements and a classical system-level network simulation setup. Although the proposed method does only require a fraction of the computation time of the latter, it achieves a significantly better match with the real world evaluations.