This thesis investigates machine learning techniques as powerful tools for the analysis and the optimal management of mobile networks towards Pervasive Intelligence. In this context, it describes recent solutions conceived for addressing three different, but related, issues. Firstly, after tailoring an unsupervised learning methodology to characterize radio resource utilization patterns, a Multi-Task Learning model, running directly at the edge of the network, is conceived to perform data mining from the control channel of an operative mobile network. Two configurations of neural networks, based on Undercomplete or Sequence to Sequence autoencoders, are exploited to obtain common feature representations of traffic profiles. Then, softmax and fully-connected layers are used to anticipate information on the type of traffic to be served and the radio resource utilization pattern requested by each service during its execution, respectively. Moreover, a Software-Defined Networking approach is exploited to monitor users’ mobility. Consequently, the prediction of both the distribution of users among cells and the communication and computational resources they request over different look-ahead horizons is performed through a Convolutional Long Short-Term Memory architecture. This information is used to perform anticipatory allocation and distribution of users' requests via Dynamic Programming. Hence, a Tenant-driven Radio Access Network slicing enforcement scheme based on Pervasive Intelligence is proposed to avoid the radio resources over-provisioning while saving bandwidth (i.e., the Pay for What You Get paradigm). The Infrastructure Provider exploits a convolutional autoencoder to compress the information on network resources and connectivity and share it with the Tenants. In turn, each Tenant implements a Deep Deterministic Policy Gradient algorithm to dynamically adapt bandwidth requests according to its own users' requirements. The outcomes of this algorithm are then used by the Infrastructure Provider to effectively enforce network slicing. The investigation in real scenarios and the comparison against conventional approaches adopted for the analysis and the optimal management of mobile networks demonstrate the effectiveness of the proposed machine learning-based solutions. In terms of applicability, the conceived methodologies are also in line with the evolution of mobile networks, where Artificial Intelligence will be natively and pervasively integrated for enabling full network automation.
Questa tesi indaga le tecniche di machine learning come potenti strumenti per l'analisi e la gestione ottima delle reti radiomobili per ottenere l’Intelligenza Pervasiva. In questo contesto, descrive soluzioni recenti progettate per affrontare tre problemi diversi, ma correlati. In primo luogo, dopo aver adattato una metodologia di unsupervised learning per caratterizzare i pattern di utilizzo delle risorse radio, un modello di multi-task learning, eseguito direttamente ai bordi della rete, è ideato per eseguire il data mining dal canale di controllo di una rete radiomobile operativa. Due configurazioni di reti neurali, basate su autoencoder Undercomplete o Sequence to Sequence, vengono sfruttate per ottenere rappresentazioni di feature comuni dei profili di traffico. In seguito, i softmax e fully-connected layer vengono utilizzati rispettivamente per anticipare le informazioni sul tipo di traffico da servire e il pattern di utilizzo delle risorse radio richiesto da ciascun servizio durante la sua esecuzione. Inoltre, viene sfruttato un approccio Software-Defined Networking per monitorare la mobilità degli utenti. Di conseguenza, la predizione sia della distribuzione degli utenti tra le celle che delle risorse di comunicazione e di computazione richieste su diversi orizzonti temporali futuri viene eseguita attraverso un'architettura Convolutional Long Short-Term Memory. Queste informazioni vengono utilizzate per eseguire in anticipo l'allocazione e la distribuzione delle richieste degli utenti tramite Dynamic Programming. Infine, viene proposto uno schema Tenant-driven di slicing enforcement nella rete di accesso radio basato sull'Intelligenza Pervasiva per evitare l'over-provisioning delle risorse radio risparmiando la banda (ovvero, il paradigma Pay for What You Get). L’Infrastructure Provider utilizza un autoencoder convoluzionale per comprimere le informazioni sulle risorse di rete e sulla connettività e le condivide con i Tenant. A sua volta, ogni Tenant implementa un algoritmo Deep Deterministic Policy Gradient per adattare dinamicamente le richieste di banda in base alle esigenze dei propri utenti. I risultati di questo algoritmo vengono quindi utilizzati dall’Infrastructure Provider per applicare efficacemente lo slicing della rete. L'indagine in scenari reali e il confronto con approcci convenzionali adottati per l'analisi e la gestione ottima delle reti radiomobili dimostrano l'efficacia delle soluzioni proposte basate sul machine learning. In termini di applicabilità, le metodologie progettate sono anche in linea con l'evoluzione delle reti radiomobili, in cui l'Intelligenza Artificiale sarà integrata in modo nativo e pervasivo per consentire la piena automazione della rete.
Analysis and Optimal Management of Mobile Networks Towards Pervasive Intelligence
Rago, Arcangela
2022
Abstract
This thesis investigates machine learning techniques as powerful tools for the analysis and the optimal management of mobile networks towards Pervasive Intelligence. In this context, it describes recent solutions conceived for addressing three different, but related, issues. Firstly, after tailoring an unsupervised learning methodology to characterize radio resource utilization patterns, a Multi-Task Learning model, running directly at the edge of the network, is conceived to perform data mining from the control channel of an operative mobile network. Two configurations of neural networks, based on Undercomplete or Sequence to Sequence autoencoders, are exploited to obtain common feature representations of traffic profiles. Then, softmax and fully-connected layers are used to anticipate information on the type of traffic to be served and the radio resource utilization pattern requested by each service during its execution, respectively. Moreover, a Software-Defined Networking approach is exploited to monitor users’ mobility. Consequently, the prediction of both the distribution of users among cells and the communication and computational resources they request over different look-ahead horizons is performed through a Convolutional Long Short-Term Memory architecture. This information is used to perform anticipatory allocation and distribution of users' requests via Dynamic Programming. Hence, a Tenant-driven Radio Access Network slicing enforcement scheme based on Pervasive Intelligence is proposed to avoid the radio resources over-provisioning while saving bandwidth (i.e., the Pay for What You Get paradigm). The Infrastructure Provider exploits a convolutional autoencoder to compress the information on network resources and connectivity and share it with the Tenants. In turn, each Tenant implements a Deep Deterministic Policy Gradient algorithm to dynamically adapt bandwidth requests according to its own users' requirements. The outcomes of this algorithm are then used by the Infrastructure Provider to effectively enforce network slicing. The investigation in real scenarios and the comparison against conventional approaches adopted for the analysis and the optimal management of mobile networks demonstrate the effectiveness of the proposed machine learning-based solutions. In terms of applicability, the conceived methodologies are also in line with the evolution of mobile networks, where Artificial Intelligence will be natively and pervasively integrated for enabling full network automation.File | Dimensione | Formato | |
---|---|---|---|
34 ciclo-RAGO Arcangela.pdf
accesso aperto
Dimensione
12.72 MB
Formato
Adobe PDF
|
12.72 MB | Adobe PDF | Visualizza/Apri |
I documenti in UNITESI sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/20.500.14242/64428
URN:NBN:IT:POLIBA-64428