Time Series Analysis to Predict Link Quality of Wireless Community Networks

Identificador:  imarina:5129027

Handle: https://hdl.handle.net/20.500.11797/imarina5129027

Autors:  Millan, Pere; Molina, Carlos; Medina, Esunly; Vega, Davide; Meseguer, Roc; Braem, Bart; Blondia, Chris

Resum:
© 2015 Elsevier B.V. Allrightsreserved. Community networks have emerged under the mottos break the strings that are limiting you, don't buy the network, be the network or a free net for everyone is possible. Such networks create a measurable social impact as they provide to the community the right and opportunity of communication. As any other network that mixes wired and wireless links, the routing protocol must face several challenges that arise from the unreliable nature of the wireless medium. Link quality tracking helps the routing layer to select links that maximize the delivery rate and minimize traffic congestion. Moreover, link quality prediction has proved to be a technique that surpasses link quality tracking by foreseeing which links are more likely to change its quality. In this work, we focus on link quality prediction by means of a time series analysis. We apply this prediction technique in the routing layer of large-scale, distributed, and decentralized networks. We demonstrate that it is possible to accurately predict the link quality in 98% of the instances, both in the short and the long terms. Particularly, we analyse the behaviour of the links globally to identify the best prediction algorithm and metric, the impact of lag windows in the results, the prediction accuracy some time steps ahead into the future, the degradation of prediction over time, and the correlation of prediction with topological features. Moreover, we also analyse the behaviour of links individually to identify the variability of link quality prediction between links, and the variability of link quality prediction over time. Finally, we also present an optimized prediction method that considers the knowledge about the expected link quality values.