Large-scale power failures induced by severe weather have become frequent and damaging in recent years, causing millions of people to be without electricity service for days. Although the power industry has been battling weather-induced failures for years, it is largely unknown how resilient the energy infrastructure and services really are to severe weather disruptions. What fundamental issues govern the resilience? Can advanced approaches such as modeling and data analytics help industry to go beyond empirical methods? This paper discusses the research to date and open issues related to these questions. The focus is on identifying fundamental challenges and advanced approaches for quantifying resilience. In particular, the first aspect of this problem is how to model large-scale failures, recoveries, and impacts, involving the infrastructure, service providers, customers, and weather. The second aspect is how to identify generic vulnerability in the infrastructure and services through large-scale data analytics. The third aspect is to understand what resilience metrics are needed and how to develop them.

Smart energy grid is an emerging area for new applications of machine learning in a nonstationary environment. Such a nonstationary environment emerges when large-scale failures occur at power networks because of external disruptions such as hurricanes and severe storms. Power distribution networks lie at the edge of the grid, and are especially vulnerable to external disruptions. Quantifiable approaches are lacking and needed to learn nonstationary behaviors of large-scale failure and recovery of power distribution. This paper studies such nonstationary behaviors in three aspects. First, a novel formulation is derived for an entire life cycle of large-scale failure and recovery of power distribution. Second, spatial-temporal models of failure and recovery of power distribution are developed as geolocation-based multivariate nonstationary GI(t)/G(t)/∞ queues. Third, the nonstationary spatial-temporal models identify a small number of parameters to be learned. Learning is applied to two real-life examples of large-scale disruptions. One is from Hurricane Ike, where data from an operational network is exact on failures and recoveries. The other is from Hurricane Sandy, where aggregated data is used for inferring failure and recovery processes at one of the impacted areas. Model parameters are learned using real data. Two findings emerge as results of learning: 1) failure rates behave similarly at the two different provider networks for two different hurricanes but differently at the geographical regions and 2) both the rapid and slow-recovery are present for Hurricane Ike but only slow recovery is shown for a regional distribution network from Hurricane Sandy.

In this work, we study cascaded disruptions between communication networks and power, using heterogeneous measurements and learning. We define sustainability of communication using distributed energy sources upon power outages, and estimate the sustainability for example networks.

Large-scale service disruptions in communication have been observed in the past but are not well-understood. The goal of this work is to gain a better understanding of disruptions in communication services in response to large-scale external disturbances such as hurricanes. In particular, Hurricane Ike is drawn as a case study, and heterogeneous data is obtained from networks, storm, and system administrators. Using the data, we first study network-wide disruptions and dependences among different unreachable subnets. Our findings show that 120 out of 230 subnets in our data set were unreachable, among which 88 subnets became unreachable dependently at a time scale of seconds or less than three minutes. We then study dependencies between communication service-disruptions and external factors such as weather and power. Unreachable subnets are found to be weakly correlated with the storm. Power outages and lack of spare power are reported to be certain causes of communication disruptions. New research issues emerge for information acquisition across communication and power infrastructures as well as weather, and information sharing among organizations.

This work provides an information-theoretic view to better understand the relationships between aggregated vulnerability information viewed by attackers and a class of randomized epidemic scanning algorithms. In particular, this work investigates three aspects: 1) anetworkvulnerabilityas the nonuniform vulnerable-host distribution, 2)threats, i.e., intelligent malwares that exploit such a vulnerability, and 3)defense, i.e., challenges for fighting the threats. We first study five large data sets and observe consistent clustered vulnerable-host distributions. We then present a new metric, referred to as thenonuniformityfactor, that quantifies the unevenness of a vulnerable-host distribution. This metric is essentially the Renyi information entropy that unifies the nonuniformity of a vulnerable-host distribution with different malware-scanning methods. Next, we draw a relationship between Renyi entropies and randomized epidemic scanning algorithms. We find that the infection rates of malware-scanning methods are characterized by the Renyi entropies that relate to the information bits in a nonunform vulnerable-host distribution extracted by a randomized scanning algorithm. Meanwhile, we show that a representative network-aware malware can increase the spreading speed by exactly or nearly a nonuniformity factor when compared to a random-scanning malware at an early stage of malware propagation. This quantifies that how much more rapidly the Internet can be infected at the early stage when a malware exploits an uneven vulnerable-host distribution as a network-wide vulnerability. Furthermore, we analyze the effectiveness of defense strategies on the spread of network-aware malwares. Our results demonstrate that counteracting network-aware malwares is a significant challenge for the strategies that include host-based defenses and IPv6.

This paper presents a large scale longitudinal study of the spatial and temporal features of malicious source addresses. The basis of our study is a 402-day trace of over 7 billion Internet intrusion attempts provided by DShield.org, which includes 160 million unique source addresses. Specifically, we focus on spatial distributions and temporal characteristics of malicious sources. First, we find that one out of 27 hosts is potentially a scanning source among 232IPv4 addresses. We then show that malicious sources have a persistent, non-uniform spatial distribution. That is, more than 80% of the sources send packets from the same 20% of the IPv4 address space over time. We also find that 7.3% of malicious source addresses are unroutable, and that some source addresses are correlated. Next, we show that most sources have a short lifetime. 57.9 % of the source addresses appear only once in the trace, and 90% of source addresses appear less than 5 times. These results have implications for both attacks and defenses.

In this paper we investigate the fault diagnosis problem in IP networks. We provide a lower bound on the average number of probes per edge using variational inference technique proposed in the context of graphical models under noisy probe measurements. To obtain the bounds, we construct a graphical model using Bayesian networks. The advantages of the variational inference technique are the explicit choices of a simplifying conjugate function and a computationally tolerable approximation to address the intractable detection problem for large networks. We propose an entropy lower (EL) bound by drawing similarities between the coding problem over binary symmetric channel and the diagnosis problem and compare it against the variational lower bound. In addition, we discuss scalable and non-scalable scenarios in the presence of noise. Simulation results demonstrate that indeed the variational inference technique can provide a linear growth of the average number of probes per edge as a function of the network size.

This work investigates three aspects: (a) a network vulnerability as the non-uniform vulnerable-host distribution, (b) threats, i.e., intelligent worms that exploit such a vulnerability, and (c) defense, i.e., challenges for fighting the threats. We first study five data sets and observe consistent clustered vulnerable-host distributions. We then present a new metric, referred to as the non-uniformity factor, which quantifies the unevenness of a vulnerable-host distribution. This metric is essentially the Renyi information entropy and better characterizes the non-uniformity of a distribution than the Shannon entropy. We then analytically and empirically measure the infection rate and the propagation speed of network-aware worms. We show that a representative network-aware worm can increase the spreading speed by exactly or nearly a non-uniformity factor when compared to a random-scanning worm at the early stage of worm propagation. This implies that when a worm exploits an uneven vulnerable-host distribution as a network-wide vulnerability, the Internet can be infected much more rapidly. Furthermore, we analyze the effectiveness of defense strategies on the spread of network-aware worms. Our results demonstrate that counteracting network-aware worms is a significant challenge for the strategies that include host-based defense and IPv6.

Localized scanning is a simple technique used by attackers to search for vulnerable hosts. Localized scanning trades off between the local and the global search of vulnerable hosts and has been used by Code Red II and Ninida worms. As such a strategy is so simple yet effective in attacking the Internet, it is important that defenders understand the spreading ability and behaviors of localized-scanning worms. In this work, we first characterize the relationships between vulnerable-host distributions and the spread of localized-scanning worms through mathematical modeling and analysis, and compare random scanning with localized scanning. We then design an optimal localized-scanning strategy, which provides an upper bound on the spreading speed of localized-scanning self-propagating codes. Furthermore, we construct three variants of localized scanning. Specifically, the feedback localized scanning and the ping-pong localized scanning adapt the scanning methods based on the feedback from the probed host, and thus spread faster than the original localized scanning and meanwhile have a smaller variance.

An important question for secure all-optical networks (AONs) is how to incorporate security against attacks in the design and engineering of network architectures. In this work, we study the resilience of AON architectures under in-band crosstalk attacks. Crosstalk attack propagation depends on both optical devices at the physical layer and wavelength usage at the network layer. This motivates us to employ probabilistic graphical models to model attack propagation. At the physical layer, we use a directed probabilistic graph (Bayesian Belief Network) to model the attack propagation under static network traffic and a given source of attack. At the network layer, we use an undirected probabilistic graph to represent the probability distribution of active connections in the network. The cross-layer model is obtained by combining the physical- and the network-layer models into a factor graph representation. Graphical models provide an explicit representation of interactions between the physical- and the network layer. Furthermore, graphical models facilitate derivations of analytical results on resilience with respect to physical-layer vulnerability, physical topology, and network load. Specifically, we derive bounds on the network resilience for regular topologies. For ring, star, and mesh-torus networks with link-shortest path routing and all-to-all traffic, we show that the average network resilience loss grows linearly with respect to the network load when the network load is small, and polynomially with respect to the probability of attack propagation from node to node along the attacker's route. In addition, numerical results suggest that the sum-product algorithm based on the factor graph representation can be used for computationally efficient evaluation of network resilience for irregular/large topologies

In this paper we consider the theoretical and practical management resources needed to reliably localize congested nodes/links and propose a solution based on decoding linear error control codes. We define the scalability of measurement-based network monitoring and relate this to the problem decoding of linear error control codes on a binary symmetric channel. Our main goal is to minimize the number of probe packets required to fully identify congested nodes and our approach requires a number of measurements that grows linearly with respect to the network size, i.e. the approach is scalable. We provide fundamental limits on the relationship between the number of probe packets, size of the network and the ability to perfectly identify all congested nodes. To identify the congested nodes we construct a bipartite graph, and congestion is inferred using the message-passing algorithm. Simulation results demonstrate the ability to perfectly locate congested nodes using this approach.

We study the resilience of all-optical network (AON) architectures under in-band crosstalk attacks. We first develop a cross-layer model that captures attack propagation based on probabilistic graphical models. At the physical layer, we use a directed probabilistic graph (Bayesian belief network) to model the attack propagation under static network traffic and a given source of attack. At the network layer, we use an undirected probabilistic graph (random field) to represent the probability distribution of active connections in the network. The cross-layer model is obtained by combining the physicaland the network-layer models into a factor graph representation. We then derive bounds on the network resilience for regular topologies. We show that for ring, star, and mesh-torus networks with link-shortest path routing and all-to-all traffic, the average network resilience loss grows linearly with respect to the network load when the network load is light; grows polynomially with respect to the probability of attack propagation from node to node along the attacker's route. We then show that the sum-product algorithm can be used for computationally efficient evaluation of network resilience for irregular topologies

Most Internet worms use random scanning. The distribution of vulnerable hosts on the Internet, however, is highly non-uniform over the IP-address space. This implies that random scanning wastes many scans on invulnerable addresses, and more virulent scanning schemes may take advantage of the non-uniformity of a vulnerable-host distribution. Questions then arise how attackers may make use of such information, and how virulent the resulting worm may be. These issues provide "worst-case scenarios" for defenders and "best-case scenarios" for attackers if the vulnerable-host distribution is available. This work develops such a scenario as the so-called importance scanning. Importance scanning results from importance sampling in statistics that scans IP-address space according to an empirical distribution of vulnerable hosts. An analytical model is developed to relate the infection rate of worms with the importance-scanning strategies. Experimental results based on parameters chosen from code red and slammer worms show that an importance-scanning worm can spread much faster than both a random-scanning worm and a routing worm. Furthermore, a game-theory approach suggests that the best strategy for defenders is to scatter applications uniformly in the entire IP-address space.

This work studies distributed configuration management of large wireless sensor networks, where management objectives are achieved by local cooperation of individual nodes. Specifically, we study when distributed configuration management is nearly optimal, and how to obtain a nearly-optimal configuration through decentralized adaptation. We first derive a spatial network model that is determined by internal network characteristics and management requirements. We next show that a sufficient condition for distributed configuration management to be nearly-optimal is that the spatial network model belongs to a class of coupled Markov random fields also known as random-bond ising model. Such graphs possess a cross-layer spatial Markov property. We specify the sufficient conditions for the nearly-optimality under different channels and density of nodes. We derive a nearly-optimal distributed algorithm using the probabilistic inference based on the derived network model. The algorithm is applied to spatial-reuse TDMA which configures a logical topology.

Network security is an important task of network management. One threat to network security is malware (malicious software) propagation. One type of malware is called topological scanning that spreads based on topology information. The focus of this work is on modeling the spread of topological malwares, which is important for understanding their potential damages, and for developing countermeasures to protect the network infrastructure. Our model is motivated by probabilistic graphs, which have been widely investigated in machine learning. We first use a graphical representation to abstract the propagation of malwares that employ different scanning methods. We then use a spatial-temporal random process to describe the statistical dependence of malware propagation in arbitrary topologies. As the spatial dependence is particularly difficult to characterize, the problem becomes how to use simple (i.e., biased) models to approximate the spatially dependent process. In particular, we propose the independent model and the Markov model as simple approximations. We conduct both theoretical analysis and extensive simulations on large networks using both real measurements and synthesized topologies to test the performance of the proposed models. Our results show that the independent model can capture temporal dependence and detailed topology information and, thus, outperforms the previous models, whereas the Markov model incorporates a certain spatial dependence and, thus, achieves a greater accuracy in characterizing both transient and equilibrium behaviors of malware propagation.

We investigate the necessary amount of network management information for light-path assessment to dynamically set up end-to-end light-paths across administrative domains in optical networks. Our focus is on the scalability of partial management information. We pose light-path assessment as a decision problem, and define the performance as the Bayes probability of an erroneous decision. We then characterize the scalability of management information as its growth rate with respect to the total resources of the network to achieve a desired performance. Scalability is achieved if the management information needed is only a negligible fraction of the total network resources. Specifically, we consider in this work one type of partial information that grows only logarithmically with the number of wavelengths supported per link. We derive an upper bound for the Bayes error in terms of the blocking probability when a new call is presented at the entrance of the network. We evaluate the upper bound using both independent and dependent models of wavelength usage for intra- and inter-domain calls. Our study shows that there exists a "threshold effect": The Bayes error decreases to zero exponentially with respect to the load when the load is either below or above a threshold value; and is nonnegligible when the load is in a small duration around the threshold. This suggests that the partial information considered can indeed provide the desired performance, and a small percentage of erroneous decisions can be traded off to achieve a large saving in the amount of management information.

Network management information for light-path assessment to dynamically set up end-to-end lightpaths across administrative domains is investigated. Our focus is on investigating what performance can be possibly achieved given partial management information, and whether a small loss in performance can trade off with a large saving in management information. The partial information we consider includes aggregated characterization of subnetworks, and local states from wavelength converters. We cast the light-path assessment as a decision problem, and define the performance as the probability of an erroneous decision. We apply the decision theory to show that the optimal performance using the partial information is the Bayes probability of error. We derive an upper bound of the Bayes error in terms of the blocking probability. We evaluate the upper bound using both independent and dependent models of wavelength usage. Our study shows that there exits a "threshold effect": the Bayes error decreases exponentially to 0 with respect to the load when the load is either below or above a threshold value; and is nonnegligible when the load is in a small duration around the threshold. This suggests that a small percentage of error decisions can trade off with a large saving in management information.

Properties of heterogeneous network traffic have been investigated from different aspects, resulting in different understanding. Specifically, one previous work discovers that the variance of network traffic exhibits a linear relationship with respect to the mean. Such a linear relation suggests that the traffic is "Poisson-like", and thus "smooth". On the other hand, prior work has shown that the heterogeneous traffic can be long-range dependent, and is thus bursty. The focus of this work is to investigate these seemingly contradictory issues, and to provide a unified understanding on the burstiness of heterogeneous traffic. In particular, we use a simple statistic, the variance of the traffic, for our investigation. We first study variance-mean relations at a single time scale. We then investigate the behavior of variances at multiple time scales, which determines the temporal correlation structure. Finally, we provide a unified view to include most important understanding of the network traffic.

The capacity of code-division multiple-access (CDMA) networks is an important issue, especially when CDMA is used for broadband wireless networks to support high speed data traffic. In this paper, based on a CDMA cell model developed in our prior work, we develop accurate and easily calculable estimations for the capacity of CDMA over large scale fading channels. We show that user bandwidth demand, transmission power and outage requirements have significant impact on network capacity. The above factors can be traded-off with each other, which is also demonstrated by our numerical results.

Using measurements collected at network monitors to infer network conditions is a promising approach for network-centric monitoring. In this context, an important question arises: given the number and locations of network monitors, how much network management resources (e.g., the number of measurements) are needed to obtain an accurate estimate of network states? We define the scalability of measurement-based network monitoring as the growth rate of the number of measurements required for accurate network monitoring/inference with respect to the size of a network. We develop a framework for investigating the scalability in the context of multicast inference with the monitors at the edges of a network. In such a framework, network monitoring/inference can be formulated as probability density estimation of network states. The growth rate is characterized through the sample complexity, which is the number of measurements needed to accurately estimate the density. The missing data framework is introduced to estimate the growth rate, where the missing data reflect unavailable measurements at the unobservable nodes without resident monitors, and the underlying nodal packet losses. We show that when the missing information is mainly due to the number of unobservable nodes, the number of measurements needed grows linearly with the size of the network, and the measurement-based inference approach is, thus, scalable. When the missing information is mainly due to the underlying nodal packet losses, the number of measurements needed grows faster than linear with the size of the network, and the measurement-based inference approach is, thus, not scalable. Our results provide guidelines for accessing feasibility of the measurement-based inference approach, and the number of probes required. We give numerical examples to illustrate some of our results.

We investigate the CDMA capacity in the support of heterogeneous traffic. The analysis is based on a CDMA cell model developed in our prior work, which provides a comprehensive yet simple foundation for theoretical analysis. Then, we develop accurate and easily calculable estimations for the capacity of CDMA networks with traffic heterogeneity over large scale fading channels. We show that the user bandwidth demand, transmission capability and outage requirement have significant impact on network capacity. The above factors can be traded-off with each other, which is also demonstrated by our numerical results.

We study the admission control of heterogeneous traffic over the reverse link of broadband CDMA networks. With traffic heterogeneity, users can have diverse QoS requirements such as transmission rates, signal-to-interference ratio (SIR), and outage probability. On the other hand, individual users experience different levels of path loss and fading. All those factors should be considered while admitting a user. Our algorithm uses a concept called effective interference bandwidth to characterize the capability of a user to support user bandwidth demand. According to our previous work, the total bandwidth resource in a CDMA cell is divided into bandwidth available to user traffic for in-cell users, and interference bandwidth for other-cell interference. User effective interference bandwidth characterizes the maximal interference a user can tolerate, and is thus a limiting parameter of cell capacity. Basically, the cell capacity is approximately the total cell bandwidth minus the user effective interference bandwidth. Since the effective interference bandwidth term varies from user to user, each user imposes a different limit on the cell capacity. If a user with high effective interference bandwidth is admitted, it is possible that future users can be blocked, thus increasing the blocking probability. By requiring users with limited effective interference bandwidth, we optimize the overall blocking probability. In designing the admission algorithm, we first evaluate the expected average traffic in a cell. This value is used to compute the maximal allowable effective interference bandwidth. Then user effective interference bandwidth is compared with maximal allowable value for admission decision. Only users with interference bandwidth not exceeding the maximal allowable level are granted admission. Numerical results show that this scheme performs better than not checking user effective interference.

Heterogeneous network traffic possesses diverse statistical properties which include complex temporal correlation and non-Gaussian distributions. A challenge to modeling heterogeneous traffic is to develop a traffic model which can accurately characterize these statistical properties, which is computationally efficient, and which is feasible for analysis. This work develops wavelet traffic models for tackling these issues. We model the wavelet coefficients rather than the original traffic. Our approach is motivated by a discovery that although heterogeneous network traffic has the complicated short- and long-range temporal dependence, the corresponding wavelet coefficients are all "short-range" dependent. Therefore, a simple wavelet model may be able to accurately characterize complex network traffic. We first investigate what short-range dependence is important among the wavelet coefficients. We then develop the simplest wavelet model, i.e., the independent wavelet model for Gaussian traffic. We define and evaluate the (average) autocorrelation function and the buffer loss probability of the independent wavelet model for fractional Gaussian noise (FGN) traffic. This assesses the performance of the independent wavelet model, and the use of which for analysis. We also develop (low-order) Markov wavelet models to capture additional dependence among the wavelet coefficients. We show that an independent wavelet model is sufficiently accurate, and a Markov wavelet model only improves the performance marginally. We further extend the wavelet models to non-Gaussian traffic through developing a novel time-scale shaping algorithm. The algorithm is tested using real network traffic and shown to outperform FARIMA in both efficiency and accuracy. Specifically, the wavelet models are parsimonious, and have a computational complexity O(N) in developing a model from a training sequence of length N, and O(M) in generating a synthetic traffic trace of length M.

Cheng Ma and Chuanyi Ji (1998) and Chuanyi Ji et al. (1999) showed empirically that independent (Haar) wavelet models were parsimonious, computationally efficient and accurate in modeling heterogeneous network traffic measured by both auto-covariance functions and buffer loss rate. We also proved analytically that such models were capable of capturing any decay rate of auto-covariance functions at large lags. In this work, we focus on comparing independent (Haar) wavelet models against independent wavelet models with higher vanishing moments. It is shown that as the vanishing moments increase, the independent wavelet models have better performance in approximating the auto-covariance function at small lags. The computational cost for all independent wavelet models is O(N) for a trace of length N.

As CDMA is to be used for the third generation wireless networks to provide integrated services, it is important to study how the quality of service, such as packet loss and delay, should be guaranteed. To accomplish this, we develop a cell-wise model in which user's resource consumption is characterized, and the constraint between system bandwidth and resource consumptions of users is given. This model incorporates important factors that affect user's quality of service, including the aggregation of user traffic, user the power constraint, path loss and interference. Based on this model, we develop an admission control procedure. We further study transmission schemes using one and two channels. Numerical results show the two-channel scheme can have more users admitted, especially under heavy interference from other cells and when users' traffic is bursty.

Network fault management is a difficult task due to the lack of accurate definitions for the different types of faults. In this work we attempt to characterize network faults by studying their effects on the statistics of network traffic. The evolution of networks in terms of size and variety of network components and services makes it difficult to understand the dynamics of the traffic. Instead, by describing network fault behavior in terms of the changes they effect on the statistics of network traffic, we can develop algorithms to detect or predict network faults. The advantage is the independence of fault descriptions from symptom-specific feature vectors and the availability of historical data.

Prediction of network failures at the router level has been achieved using an intelligent agent. The intelligent agent gathers relevant MIB data from the router and provides temporally and spatially correlated predictive alarms. The time correlated abnormal changes in the individual MIB variables are spatially correlated using a novel combining scheme. The agent was implemented on a real network and seven out of nine faults were predicted. Two typical case studies are presented. The prediction time was of the order of minutes.

In our previous work, we showed empirically that independent wavelet models were parsimonious, computationally efficient, and accurate in modeling heterogeneous network traffic measured by both auto-covariance functions and buffer loss rate. In this work, we focus on auto-covariance functions, to establish a theory of independent wavelet models as unified models for heterogeneous network traffic. We have developed the theory on the approximation capability of independent wavelet models for heterogeneous traffic in terms of the decay rate of auto-covariance functions at large lags. Average auto-covariance functions of independent wavelet models have been derived and shown to be linear combinations of basis functions. Through a simple analytical expression, we have shown that the decay rate of the auto-covariance functions of independent wavelet models is determined explicitly through a single quantity called the rate function of variances of wavelet coefficients. By specifying analytical forms of the rate function, independent wavelet models have been shown as unified models of heterogeneous traffic in terms of auto-covariance functions. The simplicity of the theory thereby provides both quantitative and qualitative explanations why independent wavelet models are unified models of heterogeneous traffic.

In this work, we investigate the performance of distributed admission control with QoS provisioning and dynamical channel allocation for mobile/wireless networks. We first provide a QoS metric feasible for admission control with dynamically allocated channels. We then derive a criterion analytically using the QoS measure for distributed call admission control with dynamic channel allocation. When maximum packing is used as the dynamic channel allocation scheme, the results obtained are independent of any particular algorithm which implements dynamic channel assignments. Our results thereby provide the optimal performance achievable for the distributed admission control with the QoS provisioning by the best dynamic channel allocation scheme in the given setting.

This paper reviews recent advances in supervised learning with a focus on two most important issues: performance and efficiency. Performance addresses the generalization capability of a learning machine on randomly chosen samples that are not included in a training set. Efficiency deals with the complexity of a learning machine in both space and time. As these two issues are general to various learning machines and learning approaches, we focus on a special type of adaptive learning systems with a neural architecture. We discuss four types of learning approaches: training an individual model; combinations of several well-trained models; combinations of many weak models; and evolutionary computation of models. We explore advantages and weaknesses of each approach and their interrelations, and we pose open questions for possible future research.

We establish a wavelet model of video traffic. We model the wavelet coefficients in the wavelet domain, which is different from the existing methods which model the video traffic in the time domain. The strength of the wavelet model includes (1) a unified approach to model both the long-range and the short-range dependence in the video traffic simultaneously, (2) a computationally efficient method of developing the model and generating high quality video traffic, and (3) the feasibility of performance analysis using the model.

We investigate the performance of distributed admission control with QoS provisioning and dynamical channel allocation for mobile/wireless networks. We first provide a QoS metric feasible for admission control with dynamically allocated channels. We then derive analytically a criterion using the QoS measure for distributed call admission control with dynamic channel allocation. Since maximum packing is used as the dynamic channel allocation scheme, the results obtained are independent of any particular algorithm which implements the dynamic channel assignment. Our results thereby provides the optimal performance achievable for the distributed admission control with the QoS provisioning by the best dynamic channel allocation scheme in the given setting.

A significant discovery from this work is that although video traffic has complicated short- and long-range dependence in the time domain, the corresponding wavelet coefficients are no longer long-range dependent in the wavelet domain. Therefore, a "short-range" dependent process can be used to model video traffic in the wavelet domain. In this work, we develop such wavelet models for VBR video traffic. The strength of the developed wavelet models includes: (1) it provides a unified approach to model both long-range and short-range dependence in video traffic simultaneously, (2) it has the ability to reduce the temporal dependence so significantly that the wavelet coefficients can be modeled by either independent or Markov models, and (3) the model results in a computationally efficient method on generating high quality video traffic.

The lower and upper bounds for the information capacity of two-layer feedforward neural networks with binary interconnections, integer thresholds for the hidden units, and zero threshold for the output unit is obtained through two steps. First, through a constructive approach based on statistical analysis, it is shown that a specifically constructed (N-2L-1) network with N input units, 2L hidden units, and one output unit is capable of implementing, with almost probability one, any dichotomy of O(W/ln W) random samples drawn from some continuous distributions, where W is the total number of weights of the network. This quantity is then used as a lower bound for the information capacity C of all (N-2L-1) networks with binary weights. Second, an upper bound is obtained and shown to be O(W) by a simple counting argument. Therefore, we have /spl Omega/(W/ln W)/spl les/C/spl les/O(W).

In this work, a probabilistic model is established for recurrent networks. The expectation-maximization (EM) algorithm is then applied to derive a new fast training algorithm for recurrent networks through mean-field approximation. This new algorithm converts training a complicated recurrent network into training an array of individual feedforward neurons. These neurons are then trained via a linear weighted regression algorithm. The training time has been improved by five to 15 times on benchmark problems.

The increasing role of communication networks in today's society results in a demand for higher levels of network availability and reliability. At the same time, fault management is becoming more difficult due to the dynamic nature and heterogeneity of networks. We propose an intelligent monitoring system using adaptive statistical techniques. The system continually learns the normal behavior of the network and detects deviations from the norm. Within the monitoring system, the measurements are segmented, and features extracted from the segments are used to describe the normal behavior of the measurement variables. This information is combined in the structure of a Bayesian network. The proposed system is thereby able to detect unknown or unseen faults. Experimental results on real network data demonstrate that the proposed system can detect abnormal behavior before a fault actually occurs.

Generalization is one of the most important problems in neural-network research. It is influenced by several factors in the network design, such as network size, weight decay factor, and others. We show here that the initial weight distribution (for gradient decent training algorithms) is one other factor that influences generalization. The initial conditions guide the training algorithm to search particular places of the weight space. For instance small initial weights tend to result in low complexity networks, and therefore can effectively act as a regularization factor. We propose a novel network complexity measure, which is helpful in shedding insight into the phenomenon, as well as in studying other aspects of generalization.

To obtain classification systems with both good generalization performance and efficiency in space and time, we propose a learning method based on combinations of weak classifiers, where weak classifiers are linear classifiers (perceptrons) which can do a little better than making random guesses. A randomized algorithm is proposed to find the weak classifiers. They are then combined through a majority vote. As demonstrated through systematic experiments, the method developed is able to obtain combinations of weak classifiers with good generalization performance and a fast training time on a variety of test problems and real applications. Theoretical analysis on one of the test problems investigated in our experiments provides insights on when and why the proposed method works. In particular, when the strength of weak classifiers is properly chosen, combinations of weak classifiers can achieve a good generalization performance with polynomial space- and time-complexity.

We study how certain smoothness constraints, for example, piecewise continuity, can be generalized from a discrete set of analog-valued data, by modifying the error backpropagation, learning algorithm. Numerical simulations demonstrate that by imposing two heuristic objectives — (1) reducing the number of hidden units, and (2) minimizing the magnitudes of the weights in the network — during the learning process, one obtains a network with a response function that smoothly interpolates between the training data.

The beam pattern of an aperture of sensor elements determines the quality of the image of far-field sources produced by it. Resolution is limited by the physical size of the aperture in linear beamforming, and sidelobe characteristics are determined by the aperture weighting function or aperture apodization. For a given number of elements the largest two-dimensional physical apertures are formed if elements are distributed only around their boundaries. Results developed recently are extended to show that in general, for any convex boundary aperture in a plane, it is possible to synthesize any beam pattern which is obtainable from a filled aperture with the same boundary. Specific results for elliptical and rectangular boundaries are briefly reviewed, and a general synthesis theorem which provides the basis of the two-dimensional pattern synthesis scheme is established. This is applicable for both incoherent and coherent imaging, the latter implemented using a transmit-receive scheme.>