The high variability of the internet traffic and the static planning of the network resources, create overprovisioning outside the peak hours that leads to revenue decrease and network efficiency degradation. In this work we are focusing on machine learning methods that allow to build analytical models from particular inputs in order to make predictions or decisions. Using algorithms that iteratively learn from data and dynamic network operations, we are able to solve these problems. In this work we used three datasets. The first one is a set of Call Detail Records (CDRs) in which are stored information of voice/sms/data of mobile users for each squared cell of the Milan area, measured during November and December 2013. The second dataset is a set of Base Station (BS) locations information, as the type of the BS (GSM, UMTS, LTE) and GPS position. The last dataset describes the land usage of the entire Lombardy region. Machine learning methods have been applied to the first dataset to make predictions of CDRs and clustering, in which typical patterns are detected and analized in order to better manage the resources allocation of the metro network in Milan. Then, thanks to the second dataset, the network topology and the bandwidth demand were estimated with the purpose of implementing a dynamic bandwidth allocation. In the end, we used a descrete event simulator to get optimal solutions of Routing and Wavelength Assignment (RWA) by solving Integer Linear Programming (ILP) problems. The project builds a predictor and a pattern detector given the spatio-temporal fluctuations of the traffic in the mobile network. The results have been used as an input for the simulator to reproduce a metro network and optimaze dynamically the resource allocation. Thanks to the predictability of traffic, all optimization problems can be solved offline, and then the solutions found can be downloaded in the network at reconfiguration time points. Keywords: machine learning, prediction, clustering, pattern recognition.
L'elevata variabilità del traffico Internet e la statica pianificazione delle risorse di rete, creano un overprovisioning al di fuori delle ore di punta che porta alla diminuzione dei ricavi e degrada l'efficienza della rete. In questo lavoro ci concentreremo sui metodi di apprendimento automatico che permettono di costruire modelli analitici da particolari ingressi al fine di fare previsioni sul traffico o prendere decisioni. Utilizzando algoritmi che imparano iterativamente dai dati in ingresso e implementando operazioni dinamiche di pianificazione della rete, siamo in grado di risolvere questi problemi. In questo lavoro ho usato tre dataset. Il primo è un insieme di Call Detail Records (CDRs) in cui risiedono informazioni come voce/sms/traffico dati dell’utente mobile in ogni cella di Milano, misurati durante i mesi di Novembre e Dicembre 2013. Il secondo dataset contiene le informazioni riguardanti le base stations (BS) dell’intero territorio italiano. Queste informazioni riguardano il tipo (GSM,UMTS,LTE) e la posizione geografica. L’ultimo dataset descrive l’utilizzo del suolo dell’intera regione Lombardia. I metodi di machine learning sono stati applicati al primo dataset per fare predizioni dei CDR e clustering, in cui pattern ricorrenti sono stati individuati e analizzati al fine di gestire al meglio l’allocazione di risorse della rete metropolitana di Milano. Dopodiché, grazie al secondo dataset, è stato possibile stimare la topologia di rete e le domande di banda con l’obiettivo di implementare un allocazione dinamica delle risorse. Infine, ho utilizzato un simulatore ad eventi discreti al fine di ottenere le soluzioni ai problemi di assegnamento dell’instradamento delle domande e delle lunghezze d’onda. Il progetto costruisce un modello di predizione e di riconoscimento dei pattern date le fluttuazioni spazio temporali del traffico nella rete mobile. I risultati sono stati usati come input per il simulatore con l’obiettivo di simulare la rete e ottimizzare dinamicamente le risorse. Grazie alla predicibilità del traffico, tutti i problemi di ottimizzazione possono essere risolti offline, e le soluzioni trovate possono essere scaricate nella rete durante i momenti di riconfigurazione della stessa.
Machine learning-based traffic prediction and pattern extraction for dynamic optical routing in SDN mobile metro networks
TROIA, SEBASTIAN
2015/2016
Abstract
The high variability of the internet traffic and the static planning of the network resources, create overprovisioning outside the peak hours that leads to revenue decrease and network efficiency degradation. In this work we are focusing on machine learning methods that allow to build analytical models from particular inputs in order to make predictions or decisions. Using algorithms that iteratively learn from data and dynamic network operations, we are able to solve these problems. In this work we used three datasets. The first one is a set of Call Detail Records (CDRs) in which are stored information of voice/sms/data of mobile users for each squared cell of the Milan area, measured during November and December 2013. The second dataset is a set of Base Station (BS) locations information, as the type of the BS (GSM, UMTS, LTE) and GPS position. The last dataset describes the land usage of the entire Lombardy region. Machine learning methods have been applied to the first dataset to make predictions of CDRs and clustering, in which typical patterns are detected and analized in order to better manage the resources allocation of the metro network in Milan. Then, thanks to the second dataset, the network topology and the bandwidth demand were estimated with the purpose of implementing a dynamic bandwidth allocation. In the end, we used a descrete event simulator to get optimal solutions of Routing and Wavelength Assignment (RWA) by solving Integer Linear Programming (ILP) problems. The project builds a predictor and a pattern detector given the spatio-temporal fluctuations of the traffic in the mobile network. The results have been used as an input for the simulator to reproduce a metro network and optimaze dynamically the resource allocation. Thanks to the predictability of traffic, all optimization problems can be solved offline, and then the solutions found can be downloaded in the network at reconfiguration time points. Keywords: machine learning, prediction, clustering, pattern recognition.| File | Dimensione | Formato | |
|---|---|---|---|
|
thesis.pdf
accessibile in internet per tutti
Descrizione: Testo della Tesi
Dimensione
3.45 MB
Formato
Adobe PDF
|
3.45 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/126134