Dissertations / Theses on the topic 'Limited network resources'

Thesis (Ph. D.)--Massachusetts Institute of Technology, Dept. of Electrical Engineering and Computer Science, 2008.
This electronic version was submitted by the student author. The certified thesis is available in the Institute Archives and Special Collections.
Includes bibliographical references (leaves 201-207).
We consider the problem of decentralized binary detection in a network consisting of a large number of nodes arranged as a tree of bounded height. We show that the error probability decays exponentially fast with the number of nodes under both a Neyman-Pearson criterion and a Bayesian criterion, and provide bounds for the optimal error exponent. Furthermore, we show that under the Neyman-Pearson criterion, the optimal error exponent is often the same as that corresponding to a parallel configuration, implying that a large network can be designed to operate efficiently without significantly affecting the detection performance. We provide sufficient, as well as necessary, conditions for this to happen. For those networks satisfying the sufficient conditions, we propose a simple strategy that nearly achieves the optimal error exponent, and in which all non-leaf nodes need only send 1-bit messages. We also investigate the impact of node failures and unreliable communications on the detection performance. Node failures are modeled by a Galton-Watson branching process, and binary symmetric channels are assumed for the case of unreliable communications. We characterize the asymptotically optimal detection performance, develop simple strategies that nearly achieve the optimal performance, and compare the performance of the two types of networks. Our results suggest that in a large scale sensor network, it is more important to ensure that nodes can communicate reliably with each other(e.g.,by boosting the transmission power) than to ensure that nodes are robust to failures. In the case of networks with unbounded height, we establish the validity of a long-standing conjecture regarding the sub-exponential decay of Bayesian detection error probabilities in a tandem network. We also provide bounds for the error probability, and show that under the additional assumption of bounded Kullback-Leibler divergences, the error probability is (e cnd ), for all d> 1/2, with c c(logn)d being a positive constant. Furthermore, the bound (e), for all d> 1, holds under an additional mild condition on the distributions. This latter bound is shown to be tight. Moreover, for the Neyman-Pearson case, we establish that if the sensors act myopically, the Type II error probabilities also decay at a sub-exponential rate.
(cont.) Finally, we consider the problem of decentralized detection when sensors have access to side-information that affects the statistics of their measurements, and the network has an overall cost constraint. Nodes can decide whether or not to make a measurement and transmit a message to the fusion center("censoring"), and also have a choice of the transmission function. We study the tradeoff in the detection performance with the cost constraint, and also the impact of sensor cooperation and global sharing of side-information. In particular, we show that if the Type I error probability is constrained to be small, then sensor cooperation is not necessary to achieve the optimal Type II error exponent.
by Wee Peng Tay.
Ph.D.

Dans ce travail, notre contribution est double. Nous développons un cadre d’apprentissage stochastique distribué pour la recherche des équilibres de Nash dans le cas de fonctions de paiement dépendantes d’un état. La plupart des travaux existants supposent qu’une expression analytique de la récompense est disponible au niveau des noeuds. Nous considérons ici une hypothèse réaliste où les noeuds ont seulement une réalisation quantifiée de la récompense à chaque instant et développons un modèle stochastique d’apprentissage à temps discret utilisant une perturbation en sinus. Nous examinons la convergence de notre algorithme en temps discret pour une trajectoire limite définie par une équation différentielle ordinaire (ODE). Ensuite, nous effectuons une analyse de la stabilité et appliquons le schéma proposé dans un problème de commande de puissance générique dans les réseaux sans fil. Nous avons également élaboré un cadre de partage de ressources distribuées pour les réseaux –cloud– en nuage. Nous étudions la stabilité de l’évolution de l’équilibre de Nash en fonction du nombre d’utilisateurs. Dans ce scénario, nous considérons également le comportement des utilisateurs sociaux. Enfin nous avons également examiné un problème de satisfaction de la demande où chaque utilisateur a une demande propre à lui qui doit être satisfaite
As systems are becoming larger, it is becoming difficult to optimize them in a centralized manner due to insufficient backhaul connectivity and dynamical systems behavior. In this thesis, we tackle the above problem by developing a distributed strategic learning framework for seeking Nash equilibria under state dependent payoff functions. We develop a discrete time stochastic learning using sinus perturbation with the realistic assumption, that each node only has a numerical realization of the payoff at each time. We examine the convergence of our discrete time algorithm to a limiting trajectory defined by an ordinary differential equation (ODE). Finally, we conduct a stability analysis and apply the proposed scheme in a generic wireless networks. We also provide the application of these algorithms to real world resource utilization problems in wireless. Our proposed algorithm is applied to the following distributed optimization problems in wireless domain. Power control, beamforming and Bayesian density tracking in the interference channel. We also consider resource sharing problems in large scale networks (e.g. cloud networks) with a generalized fair payoff function. We formulate the problem as a strategic decision-making problem (i.e. a game). We examine the resource sharing game with finite and infinite number of players. Exploiting the aggregate structure of the payoff functions, we show that, the Nash equilibrium is not an evolutionarily stable strategy in the finite regime. Then, we introduce a myopic mean-field response where each player implements a mean-field-taking strategy. We show that such a mean-field-taking strategy is evolutionarily stable in both finite and infinite regime. We provide closed form expression of the optimal pricing that gives an efficient resource sharing policy. As the number of active players grows without bound, we show that the equilibrium strategy converges to a mean-field equilibrium and the optimal prices for resources converge to the optimal price of the mean-field game. Then, we address the demand satisfaction problem for which a necessary and sufficiency condition for satisfactory solutions is provided

La interfaz radio de los sistemas de comunicaciones móviles es normalmente considerada como la única limitación de capacidad en la red de acceso radio. Sin embargo, a medida que se van desplegando nuevas y más eficientes interfaces radio, y de que el tráfico de datos y multimedia va en aumento, existe la creciente preocupación de que la infraestructura de transporte (backhaul) de la red celular pueda convertirse en el cuello de botella en algunos escenarios. En este contexto, la tesis se centra en el desarrollo de técnicas de gestión de recursos que consideran de manera conjunta la gestión de recursos en la interfaz radio y el backhaul. Esto conduce a un nuevo paradigma donde los recursos del backhaul se consideran no sólo en la etapa de dimensionamiento, sino que además son incluidos en la problemática de gestión de recursos. Sobre esta base, el primer objetivo de la tesis consiste en evaluar los requerimientos de capacidad en las redes de acceso radio que usan IP como tecnología de transporte, de acuerdo a las recientes tendencias de la arquitectura de red. En particular, se analiza el impacto que tiene una solución de transporte basada en IP sobre la capacidad de transporte necesaria para satisfacer los requisitos de calidad de servicio en la red de acceso. La evaluación se realiza en el contexto de la red de acceso radio de UMTS, donde se proporciona una caracterización detallada de la interfaz Iub. El análisis de requerimientos de capacidad se lleva a cabo para dos diferentes escenarios: canales dedicados y canales de alta velocidad. Posteriormente, con el objetivo de aprovechar totalmente los recursos disponibles en el acceso radio y el backhaul, esta tesis propone un marco de gestión conjunta de recursos donde la idea principal consiste en incorporar las métricas de la red de transporte dentro del problema de gestión de recursos. A fin de evaluar los beneficios del marco de gestión de recursos propuesto, esta tesis se centra en la evaluación del problema de asignación de base, como estrategia para distribuir el tráfico entre las estaciones base en función de los niveles de carga tanto en la interfaz radio como en el backhaul. Este problema se analiza inicialmente considerando una red de acceso radio genérica, mediante la definición de un modelo analítico basado en cadenas de Markov. Dicho modelo permite calcular la ganancia de capacidad que puede alcanzar la estrategia de asignación de base propuesta. Posteriormente, el análisis de la estrategia propuesta se extiende considerando tecnologías específicas de acceso radio. En particular, en el contexto de redes WCDMA se desarrolla un algoritmo de asignación de base basado en simulatedannealing cuyo objetivo es maximizar una función de utilidad que refleja el grado de satisfacción de las asignaciones respecto los recursos radio y transporte. Finalmente, esta tesis aborda el diseño y evaluación de un algoritmo de asignación de base para los futuros sistemas de banda ancha basados en OFDMA. En este caso, el problema de asignación de base se modela como un problema de optimización mediante el uso de un marco de funciones de utilidad y funciones de coste de recursos. El problema planteado, que considera que existen restricciones de recursos tanto en la interfaz radio como en el backhaul, es mapeado a un problema de optimización conocido como Multiple-Choice Multidimensional Knapsack Problem (MMKP). Posteriormente, se desarrolla un algoritmo de asignación de base heurístico, el cual es evaluado y comparado con esquemas de asignación basados exclusivamente en criterios radio. El algoritmo concebido se basa en el uso de los multiplicadores de Lagrange y está diseñado para aprovechar de manera simultánea el balanceo de carga en la intefaz radio y el backhaul.

Thesis (S.M. in Technology and Policy)--Massachusetts Institute of Technology, Engineering Systems Division, Technology and Policy Program, 2009.
This electronic version was submitted by the student author. The certified thesis is available in the Institute Archives and Special Collections.
Cataloged from student submitted PDF version of thesis.
Includes bibliographical references (p. 62-64).
Billions of dollars are spent to develop drugs for infectious diseases in developing countries. How will these drugs along with clinical services be delivered to the patients who currently do not have access to them? Health franchises have been around since early 1990s, creating networks of shops and clinics that provide specialized care to low income individuals. This thesis attempts to understand the underlying mechanisms of successful health franchises. Two cases are taken into consideration, CFWshops in Kenya and Mi Farmacita Nacional (MFN) in Mexico. Both are pharmaceutical shops with small clinics attached to them. The two cases were examined through a framework derived from successful commercial franchises and franchise theory. The elements that were addressed include operational structure, marketing strategy, product and service offerings, monitoring of businesses, and financial structure. CFWshops and MFN had some stark differences in how they addressed each of these elements. Unlike typical commercial franchises, health franchises aim to provide social benefits to the population. This goal requires franchises to not only create a business strategy to be financially sustainable and take advantage of networks, but also show health improvements in the community. The success of a health franchise is dependent on the health impacts it provides because its mission is not to generate a profit for the stakeholders but rather the value added to the customer by providing access that was not there before.
(cont.) The comparative case analysis suggests several key recommendations. Health innovations in resource limited settings should create networks with other public and private health groups to leverage existing knowledge and best practices. This reduces cost and time of learning and allows businesses to utilize existing channels to provide access for drugs and services to individuals who currently are not receiving them.
by Ramya Sankar.
S.M.in Technology and Policy

In economical and sustainable operation of cyber-physical systems, a number of entities need to often cooperate over a communication network to solve optimization problems. A challenging aspect in the design of robust distributed solution algorithms to these optimization problems is that as technology advances and the networks grow larger, the communication bandwidth used to coordinate the solution is limited. Moreover, even though most research has focused distributed convex optimization, in cyberphysical systems nonconvex problems are often encountered, e.g., localization in wireless sensor networks and optimal power flow in smart grids, the solution of which poses major technical difficulties. Motivated by these challenges this thesis investigates distributed optimization with emphasis on limited communication for both convex and nonconvex structured problems. In particular, the thesis consists of four articles as summarized below. The first two papers investigate the convergence of distributed gradient solution methods for the resource allocation optimization problem, where gradient information is communicated at every iteration, using limited communication. In particular, the first paper investigates how distributed dual descent methods can perform demand-response in power networks by using one-way communication. To achieve the one-way communication, the power supplier first broadcasts a coordination signal to the users and then updates the coordination signal by using physical measurements related to the aggregated power usage. Since the users do not communicate back to the supplier, but instead they only take a measurable action, it is essential that the algorithm remains primal feasible at every iteration to avoid blackouts. The paper demonstrates how such blackouts can be avoided by appropriately choosing the algorithm parameters. Moreover, the convergence rate of the algorithm is investigated. The second paper builds on the work of the first paper and considers more general resource allocation problem with multiple resources. In particular, a general class of quantized gradient methods are studied where the gradient direction is approximated by a finite quantization set. Necessary and sufficient conditions on the quantization set are provided to guarantee the ability of these methods to solve a large class of dual problems. A lower bound on the cardinality of the quantization set is provided, along with specific examples of minimal quantizations. Furthermore, convergence rate results are established that connect the fineness of the quantization and number of iterations needed to reach a predefined solution accuracy. The results provide a bound on the number of bits needed to achieve the desired accuracy of the optimal solution. The third paper investigates a particular nonconvex resource allocation problem, the Optimal Power Flow (OPF) problem, which is of central importance in the operation of power networks. An efficient novel method to address the general nonconvex OPF problem is investigated, which is based on the Alternating Direction Method of Multipliers (ADMM) combined with sequential convex approximations. The global OPF problem is decomposed into smaller problems associated to each bus of the network, the solutions of which are coordinated via a light communication protocol. Therefore, the proposed method is highly scalable. The convergence properties of the proposed algorithm are mathematically and numerically substantiated. The fourth paper builds on the third paper and investigates the convergence of distributed algorithms as in the third paper but for more general nonconvex optimization problems. In particular, two distributed solution methods, including ADMM, that combine the fast convergence properties of augmented Lagrangian-based methods with the separability properties of alternating optimization are investigated. The convergence properties of these methods are investigated and sufficient conditions under which the algorithms asymptotically reache the first order necessary conditions for optimality are established. Finally, the results are numerically illustrated on a nonconvex localization problem in wireless sensor networks. The results of this thesis advocate the promising convergence behaviour of some distributed optimization algorithms on nonconvex problems. Moreover, the results demonstrate the potential of solving convex distributed resource allocation problems using very limited communication bandwidth. Future work will consider how even more general convex and nonconvex problems can be solved using limited communication bandwidth and also study lower bounds on the bandwidth needed to solve general resource allocation optimization problems.

QC 20160203

Image classification is a memory- and compute-intensive task. It is difficult to implement high-speed image classification algorithms on resource-limited systems like FPGAs and embedded computers. Most image classification algorithms require many fixed- and/or floating-point operations and values. In this work, we explore the use of binary values to reduce the memory and compute requirements of image classification algorithms. Our objective was to implement these algorithms on resource-limited systems while maintaining comparable accuracy and high speeds. By implementing high-speed image classification algorithms on resource-limited systems like embedded computers, FPGAs, and ASICs, automated visual inspection can be performed on small low-powered systems. Industries like manufacturing, medicine, and agriculture can benefit from compact, high-speed, low-power visual inspection systems. Tasks like defect detection in manufactured products and quality sorting of harvested produce can be performed cheaper and more quickly. In this work, we present ECO Jet Features, an algorithm adapted to use binary values for visual inspection. The ECO Jet Features algorithm ran 3.7x faster than the original ECO Features algorithm on embedded computers. It also allowed the algorithm to be implemented on an FPGA, achieving 78x speedup over full-sized desktop systems, using a fraction of the power and space. We reviewed Binarized Neural Nets (BNNs), neural networks that use binary values for weights and activations. These networks are particularly well suited for FPGA implementation and we compared and contrasted various FPGA implementations found throughout the literature. Finally, we combined the deep learning methods used in BNNs with the efficiency of Jet Features to make Neural Jet Features. Neural Jet Features are binarized convolutional layers that are learned through deep learning and learn classic computer vision kernels like the Gaussian and Sobel kernels. These kernels are efficiently computed as a group and their outputs can be reused when forming output channels. They performed just as well as BNN convolutions on visual inspection tasks and are more stable when trained on small models.

The emerging technology of Cyberphysical systems consists of networked computing, sensing, and actuator devices used to monitor, connect, and control physical phenomena. In order to economically and sustainably operate Cyberphysical systems, their devices need to cooperate over a communication network to solve optimization problems. For example, in smart power grids, smart meters cooperatively optimize the grid performance, and in wireless sensor networks a number of sensors cooperate to find optimal estimators of real-world parameters. A challenging aspect in the design of distributed solution algorithms to these optimization problems is that while the technology advances and the networks grow larger, the communication bandwidth available to coordinate the solution remains limited. Motivated by this challenge, this thesis investigates the convergence of distributed solution methods for resource allocation optimization problems, where gradient information is communicated at every iteration, using limited communication. This problem is approached from three different perspectives, each presented in a separate paper.  The investigation of the three papers demonstrate promises and limits of solving distributed resource allocation problems using limited communication bandwidth. Future work will consider how even more general problems can be solved using limited communication bandwidth and also study different communication constraints.

QC 20170424

To meet the exponentially increasing traffic data driven by the rapidly growing mobile subscriptions, both industry and academia are exploring the potential of a new genera- tion (5G) of wireless technologies. An important 5G goal is to achieve high data rate. Small cells with spectrum sharing and multiple-input multiple-output (MIMO) techniques are one of the most promising 5G technologies, since it enables to increase the aggregate data rate by improving the spectral efficiency, nodes density and transmission bandwidth, respectively. However, the increased interference in the densified networks will in return limit the achievable rate performance if not properly managed. The considered setup can be modeled as MIMO interference networks, which can be classified into the K-user MIMO interference channel (IC) and the K-cell MIMO interfering broadcast channel/multiple access channel (MIMO-IBC/IMAC) according to the number of mobile stations (MSs) simultaneously served by each base station (BS). The thesis considers two physical layer (PHY) resource allocation problems that deal with the interference for both models: 1) Pareto boundary computation for the achiev- able rate region in a K-user single-stream MIMO IC and 2) grouping-based interference alignment (GIA) with optimized IA-Cell assignment in a MIMO-IMAC under limited feedback. In each problem, the thesis seeks to provide a deeper understanding of the system and novel mathematical results, along with supporting numerical examples. Some of the main contributions can be summarized as follows. It is an open problem to compute the Pareto boundary of the achievable rate region for a K-user single-stream MIMO IC. The K-user single-stream MIMO IC models multiple transmitter-receiver pairs which operate over the same spectrum simultaneously. Each transmitter and each receiver is equipped with multiple antennas, and a single desired data stream is communicated in each transmitter-receiver link. The individual achievable rates of the K users form a K-dimensional achievable rate region. To find efficient operating points in the achievable rate region, the Pareto boundary computation problem, which can be formulated as a multi-objective optimization problem, needs to be solved. The thesis transforms the multi-objective optimization problem to two single-objective optimization problems–single constraint rate maximization problem and alternating rate profile optimization problem, based on the formulations of the ε-constraint optimization and the weighted Chebyshev optimization, respectively. The thesis proposes two alternating optimization algorithms to solve both single-objective optimization problems. The convergence of both algorithms is guaranteed. Also, a heuristic initialization scheme is provided for each algorithm to achieve a high-quality solution. By varying the weights in each single-objective optimization problem, numerical results show that both algorithms provide an inner bound very close to the Pareto boundary. Furthermore, the thesis also computes some key points exactly on the Pareto boundary in closed-form. A framework for interference alignment (IA) under limited feedback is proposed for a MIMO-IMAC. The MIMO-IMAC well matches the uplink scenario in cellular system, where multiple cells share their spectrum and operate simultaneously. In each cell, a BS receives the desired signals from multiple MSs within its own cell and each BS and each MS is equipped with multi-antenna. By allowing the inter-cell coordination, the thesis develops a distributed IA framework under limited feedback from three aspects: the GIA, the IA-Cell assignment and dynamic feedback bit allocation (DBA), respec- tively. Firstly, the thesis provides a complete study along with some new improvements of the GIA, which enables to compute the exact IA precoders in closed-form, based on local channel state information at the receiver (CSIR). Secondly, the concept of IA-Cell assignment is introduced and its effect on the achievable rate and degrees of freedom (DoF) performance is analyzed. Two distributed matching approaches and one centralized assignment approach are proposed to find a good IA-Cell assignment in three scenrios with different backhaul overhead. Thirdly, under limited feedback, the thesis derives an upper bound of the residual interference to noise ratio (RINR), formulates and solves a corresponding DBA problem. Finally, numerical results show that the proposed GIA with optimized IA-Cell assignment and the DBA greatly outperforms the traditional GIA algorithm.

碩士
國立成功大學
資訊工程學系碩博士班
90
In the past years, due to the popularization of broadband/wireless communication technologies, many SOHO users now construct their Intranet with low-cost embedded network devices to connect to the Internet. These low-cost network devices have low-level CPU and more limitations on system resource (such as the computation power of CPU, memory, static storage, and so on) than traditional expansive network devices. How to save resource is important for these network devices when they are applied to solve the problems of content-based packet filtering and intrusion detection. In this paper, we show the problems of existing pattern matching methods when they are implemented in a resource-limited network device. We then propose a novel multiple pattern matching method that has better performance than traditional Aho-Corasick(AC) and Boyer-Moore Horspool (BMH) algorithms and uses less resource than Set-Wise Boyer-Moore Horspool algorithm. It adopts hash table to construct an extended Boyer-Moore Hospool approach for multiple pattern matching with Set-Exclusive Shift Table. The HASH-Link-List structure of the proposed approach will have less requirement of memory resource than other Multiple Pattern Matching algorithms. The Set-Exclusive Table also helps to reduce the times of matching. The proposed pattern-matching method has been applied to content-based packet filtering.

Network coding (NC) is an optimal data dissemination technique where intermediate nodes linearly combine incoming packets. To recover a network-coded message, a sink must use a Gaussian elimination decoder, but this high-complexity decoder may not be acceptable in resource-constrained applications like sensor networks. A good alternative to Gaussian elimination is for the sink to apply the well-known belief propagation (BP) algorithm; however, the performance and complexity of BP decoding is dependent on the statistics of the linearly-combined packets. In this work, we propose two protocols that address this issue by applying fountain coding paradigms to network codes. For a two-source, single-relay, and single-sink network, named the Y-network, if the relay can network-code incoming packets while maintaining the key properties of the fountain code, then BP decoding can be applied efficiently to recover the original message. Particularly, the sink should see a Soliton-like degree distribution for efficient BP decoding. The first protocol, named Soliton-like rateless coding (SLRC), recognizes that certain encoded packets are essential for BP decoding to perform well. Therefore, the relay protects these important packets by immediately forwarding them to the sink. It can be shown analytically that the proposed scheme is resilient to nodes leaving the transmission session. Through simulations, the SLRC scheme is shown to perform better than buffer-and-forwarding, and the Distributed LT code. Although SLRC achieves good performance, the degree distribution seen by the sink is non-optimal and assumes that a large number of packets can be buffered, which may not always be possible. Extending SLRC, we propose the Improved Soliton-like Rateless Coding (ISLRC) protocol. Assuming a resource-constrained relay, the available resources at the relay are effciently utilized by performing distribution shaping; packets are intelligently linearly combined. The aggregate degree distribution for the worst case is derived and used in performing an asymptotic error analysis using an AND-OR tree analysis. Simulation results show that even under the worst case scenario of ISLRC, better performance can be achieved compared to SLRC and other existing schemes.
Thesis (Master, Electrical & Computer Engineering) -- Queen's University, 2011-10-07 21:13:03.862

Les réseaux de neurones profonds sont la pierre angulaire des systèmes à la fine pointe de la technologie pour une vaste gamme de tâches, comme la reconnaissance d'objets, la modélisation du langage et la traduction automatique. Mis à part le progrès important établi dans les architectures et les procédures de formation des réseaux de neurones profonds, deux facteurs ont été la clé du succès remarquable de l'apprentissage profond : la disponibilité de grandes quantités de données étiquetées et la puissance de calcul massive. Cette thèse par articles apporte plusieurs contributions à l'avancement de l'apprentissage profond, en particulier dans les problèmes avec très peu ou pas de données étiquetées, ou avec des ressources informatiques limitées. Le premier article aborde la question de la rareté des données dans les systèmes de recommandation, en apprenant les représentations distribuées des produits à partir des commentaires d'évaluation de produits en langage naturel. Plus précisément, nous proposons un cadre d'apprentissage multitâches dans lequel nous utilisons des méthodes basées sur les réseaux de neurones pour apprendre les représentations de produits à partir de textes de critiques de produits et de données d'évaluation. Nous démontrons que la méthode proposée peut améliorer la généralisation dans les systèmes de recommandation et atteindre une performance de pointe sur l'ensemble de données Amazon Reviews. Le deuxième article s'attaque aux défis computationnels qui existent dans l'entraînement des réseaux de neurones profonds à grande échelle. Nous proposons une nouvelle architecture de réseaux de neurones conditionnels permettant d'attribuer la capacité du réseau de façon adaptative, et donc des calculs, dans les différentes régions des entrées. Nous démontrons l'efficacité de notre modèle sur les tâches de reconnaissance visuelle où les objets d'intérêt sont localisés à la couche d'entrée, tout en maintenant une surcharge de calcul beaucoup plus faible que les architectures standards des réseaux de neurones. Le troisième article contribue au domaine de l'apprentissage non supervisé, avec l'aide du paradigme des réseaux antagoniste génératifs. Nous introduisons un cadre fléxible pour l'entraînement des réseaux antagonistes génératifs, qui non seulement assure que le générateur estime la véritable distribution des données, mais permet également au discriminateur de conserver l'information sur la densité des données à l'optimum global. Nous validons notre cadre empiriquement en montrant que le discriminateur est capable de récupérer l'énergie de la distribution des données et d'obtenir une qualité d'échantillons à la fine pointe de la technologie. Enfin, dans le quatrième article, nous nous attaquons au problème de l'apprentissage non supervisé à travers différents domaines. Nous proposons un modèle qui permet d'apprendre des transformations plusieurs à plusieurs à travers deux domaines, et ce, à partir des données non appariées. Nous validons notre approche sur plusieurs ensembles de données se rapportant à l'imagerie, et nous montrons que notre méthode peut être appliquée efficacement dans des situations d'apprentissage semi-supervisé.
Deep neural networks are the cornerstone of state-of-the-art systems for a wide range of tasks, including object recognition, language modelling and machine translation. In the last decade, research in the field of deep learning has led to numerous key advances in designing novel architectures and training algorithms for neural networks. However, most success stories in deep learning heavily relied on two main factors: the availability of large amounts of labelled data and massive computational resources. This thesis by articles makes several contributions to advancing deep learning, specifically in problems with limited or no labelled data, or with constrained computational resources. The first article addresses sparsity of labelled data that emerges in the application field of recommender systems. We propose a multi-task learning framework that leverages natural language reviews in improving recommendation. Specifically, we apply neural-network-based methods for learning representations of products from review text, while learning from rating data. We demonstrate that the proposed method can achieve state-of-the-art performance on the Amazon Reviews dataset. The second article tackles computational challenges in training large-scale deep neural networks. We propose a conditional computation network architecture which can adaptively assign its capacity, and hence computations, across different regions of the input. We demonstrate the effectiveness of our model on visual recognition tasks where objects are spatially localized within the input, while maintaining much lower computational overhead than standard network architectures. The third article contributes to the domain of unsupervised learning with the generative adversarial networks paradigm. We introduce a flexible adversarial training framework, in which not only the generator converges to the true data distribution, but also the discriminator recovers the relative density of the data at the optimum. We validate our framework empirically by showing that the discriminator is able to accurately estimate the true energy of data while obtaining state-of-the-art quality of samples. Finally, in the fourth article, we address the problem of unsupervised domain translation. We propose a model which can learn flexible, many-to-many mappings across domains from unpaired data. We validate our approach on several image datasets, and we show that it can be effectively applied in semi-supervised learning settings.

碩士
國立中正大學
電機工程研究所
105
Recently, the convolutional neural network (CNN) has been widely used in deep learning for many challenging tasks, such as image recognition. Because of complicated calculations, CNN often needs to be implemented on FPGA, GPU or ASIC to meet the performance requirement. Among these realization alternatives, FPGA has been accredited for high performance, reconfigurable, and short development. Consequently, FPGA based CNN accelerators deserve good optimization strategies in order to achieve high performance under logic, memory, and I/O bandwidth constraints. In this regard, we propose to use loop tiling and subsequently to calculate the throughput, memory bandwidth, and resource usage, all under the Roofline model. As such, we can easily find trade-off among various design parameters. Moreover, the proposed methodology can be quickly adapted to other platforms for the same purpose of prototyping CNN accelerators in FPGA.

"This dissertation takes a filtering oriented point of view and systematically addresses modeling, filtering performance analysis, and filter design issues over sensor networks with network induced constraints. The overall aim of this dissertation is to expand knowledge about theory and applications of sensor networks. More specifically, rigorous and systematic theoretical frameworks of distributed H1 filtering for estimating an unavailable state signal through noisy measurement and a disturbed plant over sensor networks are established. Furthermore, new filtering concepts, filtering theory, filter design methods, and filtering algorithms are proposed to deal with various network-induced constraints."

碩士
國立交通大學
電信工程研究所
105
Recently, the centralized management in cloud radio access network (CRAN) based small cell (SC) networks enabled by the existence of a central controller and backhaul links is developed to meet the growing data demand. The effect of load-balancing under the considerations of backhaul with limited capacity and co-channel interference is investigated in this thesis. Our main goal is to maximize energy efficiency (EE) in C-RAN based SC network through a joint decision strategy consists of user association and radio resource allocation with quality-of-service (QoS) support. In addition, the SC with no user serving can be switched into sleep mode to save the energy consumption. To tackle this mixed combinatorial optimization problem, a load-balanced user association and radio resource allocation (LBUR) mechanism based on the quantum-behaved particle swarm optimization algorithm is proposed to seek out the optimal solution for the policies of user association and subchannel assignment as well as transmit power allocation. Simulation results demonstrate that the objectives of QoS satisfaction and energy conservation can be realized via load-balancing procedure in our proposed LBUR mechanism. Furthermore, it can be found that the load-balancing effect is mainly influenced by the limitation of backhaul capacity.

碩士
國立交通大學
電信工程研究所
105
To alleviate green house effect, high network energy efficiency (EE) has increasingly become an important research target in wireless communication systems. Therefore, the investigation for subchannel and transmit power allocation to mitigate the co-tier interference in the small cell network (SCN) is provided. Moreover, triggered by the merits of cloud radio access network (C-RAN), a groups of small cell base stations (SBSs) can be decomposed of a central small cell (CSC) and remote small cells (RSCs). Given that all the RSCs can be centrally controlled by the CSC to achieve the coordination, the split medium access control (MAC)-based functional splitting is adopted for the C-RAN network with scheduler in the CSCs and hybrid automatic repeat request (HARQ) functions left in the RSCs. However, the difficulty of limited fronthaul capacity puts severe impact for the RSCs to satisfy the quality-of-service (QoS) requirements of users. As a result, the traffic control mechanism is designed in this paper to overcome above-mentioned difficulty and obtain better EE performance. The traffic control-based user association and resource allocation (TURA) algorithm is proposed for a centralized resource management of a localized SCN. Consider there is hardware limitation for a CSC, it is infeasible for a single CSC to control all the RSCs in a large scale SCN. Accordingly, this paper proposes a hybrid controlled user and resource management (HARM) scheme, where a CSC centrally performs TURA for the RSCs to mitigate intra-group interference within localized C-RANs and the CSCs among different localized C-RANs conduct a cooperative resource competition (CRC) for inter-group interference alleviation. Although the CSCs in respective groups conduct TURA distributively, the CSCs will reach the correlated equilibrium (CE) by means of their own observations and the probability of taken decision for resource assignments in proposed CRC scheme, which is adopted from the regret-based learning algorithm. Simulation results verify the effect on traffic control mechanism in TURA scheme and the convergence in CRC scheme. Moreover, the comparison of system performance between proposed TURA, HARM, and CRC schemes are also analyzed. It can be shown that the TURA scheme outperforms the other schemes without the consideration of infeasible control ability for numerous RSCs, while the proposed HARM scheme falls a little performance on EE with consideration of feasible implementation.

To meet the exponentially increasing traffic data driven by the rapidly growing mobile subscriptions, both industry and academia are exploring the potential of a new genera- tion (5G) of wireless technologies. An important 5G goal is to achieve high data rate. Small cells with spectrum sharing and multiple-input multiple-output (MIMO) techniques are one of the most promising 5G technologies, since it enables to increase the aggregate data rate by improving the spectral efficiency, nodes density and transmission bandwidth, respectively. However, the increased interference in the densified networks will in return limit the achievable rate performance if not properly managed. The considered setup can be modeled as MIMO interference networks, which can be classified into the K-user MIMO interference channel (IC) and the K-cell MIMO interfering broadcast channel/multiple access channel (MIMO-IBC/IMAC) according to the number of mobile stations (MSs) simultaneously served by each base station (BS). The thesis considers two physical layer (PHY) resource allocation problems that deal with the interference for both models: 1) Pareto boundary computation for the achiev- able rate region in a K-user single-stream MIMO IC and 2) grouping-based interference alignment (GIA) with optimized IA-Cell assignment in a MIMO-IMAC under limited feedback. In each problem, the thesis seeks to provide a deeper understanding of the system and novel mathematical results, along with supporting numerical examples. Some of the main contributions can be summarized as follows. It is an open problem to compute the Pareto boundary of the achievable rate region for a K-user single-stream MIMO IC. The K-user single-stream MIMO IC models multiple transmitter-receiver pairs which operate over the same spectrum simultaneously. Each transmitter and each receiver is equipped with multiple antennas, and a single desired data stream is communicated in each transmitter-receiver link. The individual achievable rates of the K users form a K-dimensional achievable rate region. To find efficient operating points in the achievable rate region, the Pareto boundary computation problem, which can be formulated as a multi-objective optimization problem, needs to be solved. The thesis transforms the multi-objective optimization problem to two single-objective optimization problems–single constraint rate maximization problem and alternating rate profile optimization problem, based on the formulations of the ε-constraint optimization and the weighted Chebyshev optimization, respectively. The thesis proposes two alternating optimization algorithms to solve both single-objective optimization problems. The convergence of both algorithms is guaranteed. Also, a heuristic initialization scheme is provided for each algorithm to achieve a high-quality solution. By varying the weights in each single-objective optimization problem, numerical results show that both algorithms provide an inner bound very close to the Pareto boundary. Furthermore, the thesis also computes some key points exactly on the Pareto boundary in closed-form. A framework for interference alignment (IA) under limited feedback is proposed for a MIMO-IMAC. The MIMO-IMAC well matches the uplink scenario in cellular system, where multiple cells share their spectrum and operate simultaneously. In each cell, a BS receives the desired signals from multiple MSs within its own cell and each BS and each MS is equipped with multi-antenna. By allowing the inter-cell coordination, the thesis develops a distributed IA framework under limited feedback from three aspects: the GIA, the IA-Cell assignment and dynamic feedback bit allocation (DBA), respec- tively. Firstly, the thesis provides a complete study along with some new improvements of the GIA, which enables to compute the exact IA precoders in closed-form, based on local channel state information at the receiver (CSIR). Secondly, the concept of IA-Cell assignment is introduced and its effect on the achievable rate and degrees of freedom (DoF) performance is analyzed. Two distributed matching approaches and one centralized assignment approach are proposed to find a good IA-Cell assignment in three scenrios with different backhaul overhead. Thirdly, under limited feedback, the thesis derives an upper bound of the residual interference to noise ratio (RINR), formulates and solves a corresponding DBA problem. Finally, numerical results show that the proposed GIA with optimized IA-Cell assignment and the DBA greatly outperforms the traditional GIA algorithm.

An ever increasing number of wirelessly-enabled applications places a very high demand on stringent spectral resources. Cognitive radios have the potential of enhancing spectral efficiency by improving the usage of channels that are already licensed for a specific purpose. Research on cognitive radios involves answering questions such as: how can a cognitive radio transmit at a high data rate while maintaining the same quality of service for the licensed user? There are multiple forms of cognition studied in literature, and each of these models must be studied in detail to understand its impact on the overall system performance. Specifically, the information-theoretic capacity of such systems is of great interest. Also, the design of cognitive radio is necessary to achieve those capacities in real applications. In this dissertation, we formulate different problems that relate to the performance of such systems and methods to increase their efficiency. This dissertation discusses, firstly, the means of "sensing" in cognitive systems, secondly, the optimal resource allocation algorithms for interweave cognitive radio, and finally, the fundamental limits of partially and overly cognitive overlay systems.
text

With the ever-growing demand for higher data rate in next generation communication systems, researchers are pushing the limits of the existing architecture. Due to the stochastic nature of communication channels, most systems use some form of adaptive methods to adjust the transmitting parameters and allocation of resources in order to overcome channel variations and achieve optimal throughput. We will study four cases of adaptive transmission and dynamic resource allocation in collaborative systems that are practically significant. Firstly, we study hybrid automatic repeat request (HARQ) techniques that are widely used to handle transmission failures. We propose HARQ policies that improve system throughput and are suitable for point-to-point, two-hop relay, and multi-user broadcast systems. Secondly, we study the effect of having bits of mixed SNR qualities in finite length codewords. We prove that by grouping bits according to their reliability so that each codeword contains homogeneous bit qualities, the finite blocklength capacity of the system is increased. Thirdly, we study the routing and resource allocation problem in multiple collaborative networks. We propose an algorithm that enables collaboration between networks which needs little to no side information shared across networks, but rather infers necessary information from the transmissions. The collaboration between networks provides a significant gain in overall throughput compared to selfish networks. Lastly, we present an algorithm that allocates disjoint transmission channels for our cognitive radio network in the DARPA Spectrum Collaboration Challenge (SC2). This algorithm uses the real-time spectrogram knowledge perceived by the radios and allocates channels adaptively in a crowded spectrum shared with other collaborative networks.