NeTS: Small: Meta-Modelling for Complex Engineered Networks

Project: Research project

Project Details


B Project Summary Overview. The use of complex engineered networks is pervasive in everyday life. A few examples include the internet, the power grid, and transportation networks. Yet our understanding of such networks remains limited. This proposal responds to the NSF NeTS highlighted area of meta-networking research, specifically tackling the development of scientific methods for planning and assessing experiments in order to improve our understanding of wireless networks, a type of complex engineered network. Some methods for meta-modelling include polynomial regression models, splines, and neural networks. In some form, each effectively applies screening to identify the most important among a set of factors in an experiment and to develop the meta-model, an approximation of the factor (i.e., input variables) to response (i.e., performance measures) transformation. Various assumptions and limitations underlie many of these methods for meta-modelling, including that: (1) the factors have only two levels; (2) the factors are not categorical; (3) the set of factors considered for experimentation is not too large; (4) the direction of response is known for specific factors; and (5) the data are normally distributed. To address these assumptions and limitations a new approach to screening and meta-modelling is introduced. Locating arrays (LAs) are formulated to focus on identification rather than on measurement. Consequently, they exhibit logarithmic growth in the number of factors. This makes practical the consideration of an order of magnitude more factors in experimentation. As a result, LAs have the potential to transform experimentation in huge factor spaces such as those found in complex engineered networks. LAs also address the other assumptions and limitations, though they do rely on the principle of sparsity of effects. A heavy hitters approach, similar to that used in compressive sensing, is applied in developing the metamodel; additional non-parametric techniques help mitigate the assumptions underlying the methodology. Experimentation with simulation models of wireless networks, as well as physical wireless networks such as the CREW testbed federated with GENI, is planned. The robustness of the meta-models developed (i.e., their sensitivity to the configuration of the factors) is essential to understanding the quality and usefulness for optimization, management, and control of the network. Intellectual Merit. There is a vital need for methodologies for scientific evaluation of communication networks that include the development of scientific methods for planning and assessing experiments in complex engineered networks. This proposal advances research in three primary areas: (1) The theory of LAs for objective and efficient planning of experiments. (2) The application of LAs for meta-model development in assessing experiments, with minimal assumptions. (3) The verification, validation, and robustness of the developed meta-models for various wireless networks, including both simulated and physical wireless networks. Broader Impacts. Complex engineered networks have shaped modern society. These networks have evolved into complex systems with behaviour and characteristics that cannot be characterized using traditional techniques for modelling, analysis, and design. The meta-modelling approach based on LAs works to address some of the grand challenges in wireless networks for new data-driven mathematical models. Such models facilitate quantitative evaluation of new architectures and protocols through the development of scientific methods for planning and assessing experiments in networking. This will contribute to our understanding of complex engineered networks and continue to shape modern society.
Effective start/end date10/1/143/31/20


  • National Science Foundation (NSF): $500,000.00

Fingerprint Explore the research topics touched on by this project. These labels are generated based on the underlying awards/grants. Together they form a unique fingerprint.