Distributed spatial prediction for radio environment maps reconstruction in heterogeneous wireless networks
- Autores: Vinay Prasad Chowdappa
- Directores de la Tesis: Carmen Botella Mascarell (dir. tes.)
- Lectura: En la Universitat de València ( España ) en 2017
- Idioma: español
- Tribunal Calificador de la Tesis: Ana García Armada (presid.), Gema Piñero Sipán (secret.), Tommy Svensson (voc.)
- Materias:
- Texto completo no disponible (Saber más ...)
- Resumen
Mobile data traffic is expected to grow sevenfold at a compound annual growth rate of 47 percent from 2016 to 2021. To meet these demands, wireless communication researchers and designers are turning their attention towards fith generation (5G) networks. 5G will be a key enabler for the Internet of Things (IoT), whose vision is to create an environment of everything connected everywhere and provide a platform to massive machinetype communications and mission-critical applications. Heterogeneous networks will offer ubiquitous connectivity for IoT applications through a variety of coordination and cooperation techniques. Provisioning services and supporting diverse applications requires the network to be context-aware, utilizing contextual information in real-time. 5G will have the ability to extract and process various contextual information coupled with location information to improve the overall system performance. Radio environment map (REM) is a powerful tool that leverages link contextual information and location information, to support both the traditional and disruptive technologies in addressing the challenges of 5G. Link context refers in this Thesis to the evolution of the wireless propagation channel, which can be probabilistically modeled as a multi-scale dynamical system consisting of path-loss, shadowing and small-scale fading. Since the wireless channel is location-dependent, standard regression tools can be used for channel prediction in REMs. Kriging and Gaussian process regression (GPR) are popular spatial regression tools from Geo-statistics and machine learning, respectively.
Drawbacks of Kriging and GPR are a traditional centralized prediction and their computing complexity. To address these limitations, in this Thesis, REMs are developed using a distributed incremental clustering algorithm (DICA) and distributed GPR to minimize the computational complexity of kriging and GPR, respectively. DICA is a kriging based interpolation method that employs the least number of closest measurements to leverage short range variations in the local neighborhood of the unmeasured location. Distributed GPR distributes the overall computations among the independent mobile agents. Learning and prediction phases of GPR are achieved by fi rst performing local prediction and then combining the local information using a consensus scheme to obtain a global estimate. In addition, a novel distributed learning method based on importance sampling suitable for kriging and GPR is presented. The complexity of the proposed methods is analyzed and simulation results are presented to showcase the algorithm efficacy to REM reconstruction.
