Al-Sayed, Sara (2016)
Robust Adaptation and Learning Over Networks.
Technische Universität Darmstadt
Dissertation, Erstveröffentlichung
Kurzbeschreibung (Abstract)
This doctoral dissertation centers on robust adaptive networks. Robust adaptation strategies are devised to solve typical network inference tasks such as estimation and detection in a decentralized manner in the presence of impulsive contamination. Typical in wireless communication environments, an impulsive noise process can be described as one whose realizations contain sparse, random samples of amplitude much higher than nominally accounted for. An attractive feature that these robust adaptive strategies enjoy is that neither their development nor operation hinges on the availability of exact knowledge of the noise distribution: The robust adaptive strategies are capable of learning it on-the-fly and adapting their parameters accordingly. Forgoing data fusion centers, the network agents employing these strategies rely solely on local interactions and in-network processing to perform inference tasks, which renders networks more reliable, resilient to node and link failure, scalable, and resource efficient. Distributed cooperative processing finds applications in many areas including wireless sensor networks in smart-home, environmental, and industrial monitoring; healthcare; and military surveillance.
Since adaptive systems based on the mean-square-error criterion see their performance degrade in the presence of non-Gaussian noise, the robust adaptive strategies developed in this dissertation harness nonlinear data processing and robust statistics instead to mitigate the detrimental effects of impulsive noise. To this end, a robust adaptive filtering algorithm is developed that employs an adaptive error nonlinearity. The error nonlinearity is chosen to be a convex combination of preselected basis functions where the combination coefficients are adapted jointly with the estimate of the parameter of interest such that the mean-square-error relative to the optimal error nonlinearity is minimized in each iteration.
Then, a robust diffusion adaptation algorithm of the adapt-then-combine variety is developed as an extension of its stand-alone counterpart for distributed estimation over networks where the measurements may be corrupted by impulsive noise. Each node in the network runs a combination of its neighbors’ estimates through one iteration of a local robust adaptive filter update to ameliorate the effects of contamination, leading to better overall network performance matching that of a centralized strategy at steady-state.
Finally, the robust diffusion adaptation algorithm is extended further to solve the problem of distributed detection over adaptive networks where the measurements may be corrupted by impulsive noise. The estimates generated by the robust algorithm are used as basis for the design of robust local detectors, where the form of the test- statistics and the rule for the computation of the detection thresholds are motivated by the analysis of the algorithm dynamics. Each node in the network cooperates with its neighbors, utilizing their estimates, to update its local detector. Effectively, information pertaining to the event of interest percolates across the network, leading to enhanced detection performance.
The transient and steady-state behavior of the developed algorithms are analyzed in the mean and mean-square sense using the energy conservation framework. The performance of the algorithm is also examined in the context of distributed detection. Performance is validated extensively through numerical simulations in an impulsive noise scenario, revealing the robustness of the proposed strategies in comparison with state-of-the-art algorithms as well as good agreement between theory and practice.
Typ des Eintrags: | Dissertation | ||||
---|---|---|---|---|---|
Erschienen: | 2016 | ||||
Autor(en): | Al-Sayed, Sara | ||||
Art des Eintrags: | Erstveröffentlichung | ||||
Titel: | Robust Adaptation and Learning Over Networks | ||||
Sprache: | Englisch | ||||
Referenten: | Zoubir, Prof. Abdelhak M. ; Sayed, Prof. Ali H. | ||||
Publikationsjahr: | 2 Juni 2016 | ||||
Ort: | Darmstadt | ||||
Datum der mündlichen Prüfung: | 28 April 2016 | ||||
URL / URN: | http://tuprints.ulb.tu-darmstadt.de/5493 | ||||
Kurzbeschreibung (Abstract): | This doctoral dissertation centers on robust adaptive networks. Robust adaptation strategies are devised to solve typical network inference tasks such as estimation and detection in a decentralized manner in the presence of impulsive contamination. Typical in wireless communication environments, an impulsive noise process can be described as one whose realizations contain sparse, random samples of amplitude much higher than nominally accounted for. An attractive feature that these robust adaptive strategies enjoy is that neither their development nor operation hinges on the availability of exact knowledge of the noise distribution: The robust adaptive strategies are capable of learning it on-the-fly and adapting their parameters accordingly. Forgoing data fusion centers, the network agents employing these strategies rely solely on local interactions and in-network processing to perform inference tasks, which renders networks more reliable, resilient to node and link failure, scalable, and resource efficient. Distributed cooperative processing finds applications in many areas including wireless sensor networks in smart-home, environmental, and industrial monitoring; healthcare; and military surveillance. Since adaptive systems based on the mean-square-error criterion see their performance degrade in the presence of non-Gaussian noise, the robust adaptive strategies developed in this dissertation harness nonlinear data processing and robust statistics instead to mitigate the detrimental effects of impulsive noise. To this end, a robust adaptive filtering algorithm is developed that employs an adaptive error nonlinearity. The error nonlinearity is chosen to be a convex combination of preselected basis functions where the combination coefficients are adapted jointly with the estimate of the parameter of interest such that the mean-square-error relative to the optimal error nonlinearity is minimized in each iteration. Then, a robust diffusion adaptation algorithm of the adapt-then-combine variety is developed as an extension of its stand-alone counterpart for distributed estimation over networks where the measurements may be corrupted by impulsive noise. Each node in the network runs a combination of its neighbors’ estimates through one iteration of a local robust adaptive filter update to ameliorate the effects of contamination, leading to better overall network performance matching that of a centralized strategy at steady-state. Finally, the robust diffusion adaptation algorithm is extended further to solve the problem of distributed detection over adaptive networks where the measurements may be corrupted by impulsive noise. The estimates generated by the robust algorithm are used as basis for the design of robust local detectors, where the form of the test- statistics and the rule for the computation of the detection thresholds are motivated by the analysis of the algorithm dynamics. Each node in the network cooperates with its neighbors, utilizing their estimates, to update its local detector. Effectively, information pertaining to the event of interest percolates across the network, leading to enhanced detection performance. The transient and steady-state behavior of the developed algorithms are analyzed in the mean and mean-square sense using the energy conservation framework. The performance of the algorithm is also examined in the context of distributed detection. Performance is validated extensively through numerical simulations in an impulsive noise scenario, revealing the robustness of the proposed strategies in comparison with state-of-the-art algorithms as well as good agreement between theory and practice. |
||||
Alternatives oder übersetztes Abstract: |
|
||||
URN: | urn:nbn:de:tuda-tuprints-54930 | ||||
Sachgruppe der Dewey Dezimalklassifikatin (DDC): | 600 Technik, Medizin, angewandte Wissenschaften > 620 Ingenieurwissenschaften und Maschinenbau | ||||
Fachbereich(e)/-gebiet(e): | 18 Fachbereich Elektrotechnik und Informationstechnik > Institut für Nachrichtentechnik > Signalverarbeitung 18 Fachbereich Elektrotechnik und Informationstechnik 18 Fachbereich Elektrotechnik und Informationstechnik > Institut für Nachrichtentechnik |
||||
Hinterlegungsdatum: | 19 Jun 2016 19:55 | ||||
Letzte Änderung: | 19 Jun 2016 19:55 | ||||
PPN: | |||||
Referenten: | Zoubir, Prof. Abdelhak M. ; Sayed, Prof. Ali H. | ||||
Datum der mündlichen Prüfung / Verteidigung / mdl. Prüfung: | 28 April 2016 | ||||
Export: | |||||
Suche nach Titel in: | TUfind oder in Google |
Frage zum Eintrag |
Optionen (nur für Redakteure)
Redaktionelle Details anzeigen |