#### DMCA

## Deliverable D3.4

### Citations

3867 |
Time Series Analysis
- Hamilton
- 1994
(Show Context)
Citation Context ...em to its state space representation and then use Kalman filtering for updating. Although many possible representations exist for an ARMA system, in this work the representation suggested by Hamilton =-=[16]-=- is used. Once the AR system has been converted to the state space representation, Kalman filtering can be used for updating the AR model predictions, in exactly the same way updates are done to a “pu... |

3854 |
A new approach to linear filtering and prediction problems
- Kalman
- 1960
(Show Context)
Citation Context ...S linear system equations by using as initial vector the last observation available in the training set. The second way, allowing inclusion of new data in a natural way, is by use of Kalman filtering =-=[14]-=-. 2.2.1.3 Support Vector Regression (SVR) Model Support Vector Machines (SVM) have been used for classification, regression, and time series analysis [15]. SVR intends to find a function f(y), by sol... |

3703 | Support-vector network
- Cortes, Vapnik
- 1995
(Show Context)
Citation Context .... Similarly in the context of regression, those (support) vectors within a tube of predefined margin, determine the configuration of the SVR. More technical details about SVMs and SVR can be found in =-=[30]-=- and [31]. Reservoir Computing will be tested as a supervised learning technique for predicting the class or the exact number of the bitrate increase of traffic flow aggregates (time series prediction... |

2043 |
Network flows: theory, algorithms, and applications,
- Ahuja, Magnanti, et al.
- 1993
(Show Context)
Citation Context ... a function f : V × V → R which attaches a load to every edge of the network. The function is constrained by a capacity constraint, a flow symmetry constraint and a flow conservation constraint. (see =-=[28]-=-). FP7-223936 ECODE Project - Deliverable D.3.4 - Design and Implementation of Technical Objective 2 Page 29update-distribution batch Update to central RIB/FIB Swap overhead Distribute FIB to linecar... |

1069 | A tutorial on learning with Bayesian networks.
- Heckerman, Geiger, et al.
- 1995
(Show Context)
Citation Context ... There are two distinctive phases for our proposed algorithm as follows: 1. Learning Phase: we use statistical method to train the MLE regarding SRLGs. We introduce a modified Bayesian network ([39], =-=[40]-=- ) for training. The following example describes the method in detail. We assume a simple network as shown in Fig. 4.5. Suppose a particular node observes the following LS update pattern given in Fig.... |

602 | Vivaldi: A Decentralized Network Coordinate System
- Dabek, Cox, et al.
(Show Context)
Citation Context ...ose some methods that rely on the nodes running an ICS to detect useful routing shortcuts in networks. 3.1 Problem Formalization 3.1.1 Improving the ICS We focus on a classical ICS algorithm, Vivaldi =-=[21]-=-, which approximates a network by a system of springs and seeks to minimize its energy. The minimization is fully distributed and iteratively done at each node. In each iteration, each node updates it... |

300 | The new routing algorithm for the ARPANET.
- McQuillan, Richer, et al.
- 1980
(Show Context)
Citation Context ...t path tree is usually optimized to be recalculated in its entirety and takes about 30 to 50 µs per destination prefix. Optimizations can be done using incremental SPF (iSPF) calculation schemes (see =-=[24]-=- and [25]). The second step consists of updating the central RIB and FIB, using the calculated shortest paths. This uses about 50 to 100 µs per destination prefix (see [26]). Typically this step happe... |

294 | On the constancy of Internet path properties.
- Zhang, Duffield, et al.
- 2001
(Show Context)
Citation Context ...ded into 20 minutes FP7-223936 ECODE Project - Deliverable D.3.4 - Design and Implementation of Technical Objective 2 Page 11slots. This is a reasonable assumption in relatively well connected nodes =-=[18]-=-. All the measurements that fell into a given slot were averaged. Due to network and server conditions, some of the measurements were unavailable. Indeed, if a particular server did not respond at the... |

256 | Support vector regression machines
- Drucker, Burges, et al.
- 1997
(Show Context)
Citation Context ...ly in the context of regression, those (support) vectors within a tube of predefined margin, determine the configuration of the SVR. More technical details about SVMs and SVR can be found in [30] and =-=[31]-=-. Reservoir Computing will be tested as a supervised learning technique for predicting the class or the exact number of the bitrate increase of traffic flow aggregates (time series prediction). RC pro... |

194 | Parameter estimation for linear dynamical systems,”
- Ghahramani, Hinton
- 1996
(Show Context)
Citation Context ... the dynamical evolution of the system in the hidden or dynamical space. The parameters A, B, R, Q are estimated from the collected dataset D, using an expectation maximization (EM) algorithm for LDS =-=[13]-=-. Once the model parameters have been found, there are two possible ways to predict data. The first one consists in initializing the LDS linear system equations by using as initial vector the last obs... |

189 | Predicting time series with support vector machines”, in
- Müller, Smola, et al.
- 1997
(Show Context)
Citation Context ... a natural way, is by use of Kalman filtering [14]. 2.2.1.3 Support Vector Regression (SVR) Model Support Vector Machines (SVM) have been used for classification, regression, and time series analysis =-=[15]-=-. SVR intends to find a function f(y), by solving an optimization problem, that calculates the best function fitting a regression hyperplane that passes within a ɛ distance from certain training samp... |

119 |
Self-Organization and Association Memory (3rd edition).
- Kohonen
- 1989
(Show Context)
Citation Context ...mation. A SOM is an unsupervised machine learning method often used for statistical data analysis and data clustering. This type of artificial neural network was first described by Teuvo Kohonen (see =-=[29]-=-). In essence, SOMs map the input space to a lower-dimensional space (projection on the target space) which is easier to comprehend. This mapping is made such that it preserves the statistical structu... |

110 |
Early Application Identification.
- Bernaille, Teixeira, et al.
- 2006
(Show Context)
Citation Context ...e basis for an interactive procedure which converges to a local minimum for the objective function. Clustering technique has being successfully applied for flow and network application identification =-=[44]-=-, [45], [46]. 5.2.2 Hidden Markov Model Hidden Markov Model (HMM) is a type of finite state machine having a set of initial state probabilities (π), transition probabilities (A), hidden states (Q), ou... |

105 | Flow Clustering Using Machine Learning Techniques.
- McGregor, Hall, et al.
- 2004
(Show Context)
Citation Context ...an interactive procedure which converges to a local minimum for the objective function. Clustering technique has being successfully applied for flow and network application identification [44], [45], =-=[46]-=-. 5.2.2 Hidden Markov Model Hidden Markov Model (HMM) is a type of finite state machine having a set of initial state probabilities (π), transition probabilities (A), hidden states (Q), output probabi... |

64 | New dynamic algorithms for shortest path tree computation.
- Narváez, Siu, et al.
- 2000
(Show Context)
Citation Context ...ee is usually optimized to be recalculated in its entirety and takes about 30 to 50 µs per destination prefix. Optimizations can be done using incremental SPF (iSPF) calculation schemes (see [24] and =-=[25]-=-). The second step consists of updating the central RIB and FIB, using the calculated shortest paths. This uses about 50 to 100 µs per destination prefix (see [26]). Typically this step happens in (ps... |

62 | Achieving sub-second IGP convergence in large IP networks,
- Francois, Filsfils, et al.
- 2005
(Show Context)
Citation Context ...F) calculation schemes (see [24] and [25]). The second step consists of updating the central RIB and FIB, using the calculated shortest paths. This uses about 50 to 100 µs per destination prefix (see =-=[26]-=-). Typically this step happens in (pseudo-)parallel with step 3, which is about distributing the central FIB entries towards the line cards’ LFIB. Running step 2 and 3 in (pseudo-)parallel, means that... |

54 |
Out-of-sample Tests of Forecasting Accuracy: An Analysis and Review.
- Tashman
- 2000
(Show Context)
Citation Context ...ts or samples. In order to fill the gaps, we used the K-nearest neighbors algorithm, with a window of six samples. 2.4.2 Crossvalidation A crossvalidation method called rolling window crossvalidation =-=[19]-=- is used. As defined before, the forecast origin, is the point in time n from where the predictions are going to be made, usually the last sample of the training set. The lead time is the time h for w... |

49 | T.S.E.: Towards Network Triangle Inequality Violation Aware Distributed Systems.
- Wang, Zhang, et al.
- 2007
(Show Context)
Citation Context ... ˆ dmean − ˆ dmedian d Note that the last two variables, ˆ dK d respectively. The former, ˆ dK d , ˆ dmean − d ˆdmax , ˆ dmean , ˆdstd ˆ dK d , ˆ dstd d . and ˆ dstd , are equivalent to those used in =-=[22]-=- and [4] d , is a measure of relative estimation error, denoted by REE, while the latter, ˆ dstd , is the standard deviation of the relative estimation error, denoted d by std_REE. For supervised lear... |

48 | 25 years of time series forecasting.
- Gooijer, Hyndman
- 2006
(Show Context)
Citation Context ... average of the average per lead time is found, thus the final error is obtained. 2.4.3 Error Measurements In order to measure the error made by the models, there are as much as 18 error measurements =-=[20]-=-. The Mean Absolute Percentage Error (MAPE) was selected for this work. The MAPE measures the difference in a percentual sense, between the prediction from the actual value. It is defined as, where ˆy... |

47 |
Rabiner,“ A tutorial on Hidden Markov Models and selected applications in speech recognition,”
- R
- 1989
(Show Context)
Citation Context ... learning method. Mathematical details of the HMM, demonstrating the solution of this problem and how they can be applied to be used as a supervised learning techniques can be found in the literature =-=[47]-=-, including application to network problems [48]. 5.3 Implementation The initial phase of the profile based accountability relies on simulations to validate the dual approach of classifying action pro... |

46 | A.: Traffic data repository at the wide project. In:
- Cho, Mitsuya, et al.
- 2000
(Show Context)
Citation Context ... account the given input variables such as to evaluate the resulting packet loss and needed recovery time for a given packet trace (PCAP-file) of an IP router as available by the MAWI working group ( =-=[33]-=- ). For evaluating the machine learning techniques mentioned in Section 4.1.2, MATLAB is used in combination with available toolboxes (Neural Networks, Statistics, System Identification, e.o.) and lib... |

33 |
G.M.: HMM Profiles for Network Traffic Classification. In:
- Wright, Monrose, et al.
- 2004
(Show Context)
Citation Context ...M, demonstrating the solution of this problem and how they can be applied to be used as a supervised learning techniques can be found in the literature [47], including application to network problems =-=[48]-=-. 5.3 Implementation The initial phase of the profile based accountability relies on simulations to validate the dual approach of classifying action profiles and identifying subscribers profiles (desc... |

23 | A Machine Learning approach for efficient traffic classification,
- Li, Moore
- 2008
(Show Context)
Citation Context ...s for an interactive procedure which converges to a local minimum for the objective function. Clustering technique has being successfully applied for flow and network application identification [44], =-=[45]-=-, [46]. 5.2.2 Hidden Markov Model Hidden Markov Model (HMM) is a type of finite state machine having a set of initial state probabilities (π), transition probabilities (A), hidden states (Q), output p... |

22 |
Trap avoidance and protection schemes in networks with shared risk link groups,”
- Xu
- 2003
(Show Context)
Citation Context ...inks share the same risk). Therefore, any form of cable cut for that particular physical link may manifest two simultaneous link failures and eventually the two links are part of the same SRLG ([34], =-=[35]-=-, [36]). Identification of SRLGs in the network topology reduces the failure recovery time and therefore total amount of packet losses ([34], [35], [36]). This is because, otherwise, each link failure... |

17 |
Bayesian Networks
- Ben-Gal
- 2007
(Show Context)
Citation Context ...m used There are two distinctive phases for our proposed algorithm as follows: 1. Learning Phase: we use statistical method to train the MLE regarding SRLGs. We introduce a modified Bayesian network (=-=[39]-=-, [40] ) for training. The following example describes the method in detail. We assume a simple network as shown in Fig. 4.5. Suppose a particular node observes the following LS update pattern given i... |

11 | Interdomain Traffic Engineering in a Locator/Identifier Separation Context",
- Saucez, Donnet, et al.
- 2008
(Show Context)
Citation Context ...west cost and so on. Rank value can equal cost if there is no privacy requirements (ranking can be used to hide costs in order to hide path information). An example of Cost Function might be found in =-=[17]-=-. A Cost Function can be a machine learning model that get its parameters directly in the KB. 2.4 Experimentation The PPMs under consideration are delay and throughput. Delay is a measure of the time ... |

8 |
Continuous and Discrete Signals and Systems
- Soliman, Srinath
- 1998
(Show Context)
Citation Context ...oss function. 2.2.1.2 State Space Model Linear Dynamical systems (LDS) can be described as a system of joint linear equations, describing their dynamical evolution and how these observations are made =-=[12]-=-. A state space representation of an LDS is modeled as the evolution of nonobservable variables x and observable variables y. The LDS is expressed as: yt = Axt + vt (2.3) xt = Bxt−1 + wt (2.4)... |

7 | Detecting triangle inequality violations in internet coordinate systems by supervised learning - work in progress
- Liao, Kaafar, et al.
- 2009
(Show Context)
Citation Context ...m is required. This significantly reduces the overhead of active probing and largely improves the efficiency of the network. We have used Machine Learning techniques to improve the accuracy of an ICS =-=[3, 4, 5]-=-. We have derived automatically a criterion that can be used by nodes to better select their neighbours in the ICS and thereby reduce the impact of Triangular Inequality Violations (TIVs), which are d... |

7 |
Predicting internet end-to-end delay: a statistical case study
- Yang, Ru, et al.
- 2005
(Show Context)
Citation Context ...entation of Technical Objective 2 Page 5Figure 2.1: Overview of the IDIPS service 2.1 Problem Formalization Prediction of Path Performance Metrics (PPM) is a topic that has been explored in the past =-=[8, 9, 10]-=-. However, how to adapt to the network dynamic conditions such that models reflect this is still an open question. Furthermore, how to update the models and the effects of using partial updates, that ... |

7 |
Michiel D’Haene, and Dirk Stroobandt. An experimental unification of reservoir computing methods
- Verstraeten, Schrauwen
(Show Context)
Citation Context ...nt network (the reservoir). Supervision happens by learning a (typically linear) combination of the reservoir signals to be mapped to the target values (labels). More details about RC can be found in =-=[32]-=-. 4.1.3 Implementation Input The machine learning component involved with optimizing the router update process requires the following data in order to function adequately. These data are extracted fro... |

5 | Predicting Internet end-to-end delay: A multiple-model approach
- Yang, Ru, et al.
- 2005
(Show Context)
Citation Context ...entation of Technical Objective 2 Page 5Figure 2.1: Overview of the IDIPS service 2.1 Problem Formalization Prediction of Path Performance Metrics (PPM) is a topic that has been explored in the past =-=[8, 9, 10]-=-. However, how to adapt to the network dynamic conditions such that models reflect this is still an open question. Furthermore, how to update the models and the effects of using partial updates, that ... |

5 |
A new shared-path protection algorithm under shared risk link group constraints for survivable WDM mesh networks
- Guo, Yu, et al.
- 2005
(Show Context)
Citation Context ...hare the same risk). Therefore, any form of cable cut for that particular physical link may manifest two simultaneous link failures and eventually the two links are part of the same SRLG ([34], [35], =-=[36]-=-). Identification of SRLGs in the network topology reduces the failure recovery time and therefore total amount of packet losses ([34], [35], [36]). This is because, otherwise, each link failure withi... |

4 | Overlay Routing Using Coordinate Systems
- Cantin, Gueye, et al.
- 2008
(Show Context)
Citation Context ...een nodes can also be useful to select better paths for real-time applications. We have also proposed some methods that rely on the nodes running an ICS to detect useful routing shortcuts in networks =-=[6]-=-. • Upon failures the IP router must update its routing and forwarding tables, which may take some time and lead to packet losses. We have evaluated traffic-informed router update models using strateg... |

3 | Optimizing the ip router update process with traffic-driven updates
- Tavernier, Papadimitriou, et al.
- 2009
(Show Context)
Citation Context ...were quantitatively characterized. Depending on the context, we showed that the formulated strategies can result into a decrease of packet loss of 10 to 80 percent using small router process quantums =-=[7]-=-. Because a traffic-informed router update models can only be effective if the traffic statistics that are being used are accurate, the work is being extended such as to predict the short-term trend o... |

1 | A comparative study of path performance metrics predictors
- P, Donnet, et al.
- 2009
(Show Context)
Citation Context ...oints collaborate and use clustering to estimate path performance. The paper also shows how effective is the overhead reduction and what is the impact interm of measurement accuracy. In addition, in =-=[1]-=-, we demonstrate how to reduce the path performance metric measurements by applying machine learning techniques. We express the problem as a time series regression problem and propose several adaptati... |

1 |
On the impact of clustering on measurement reduction
- D, Bonaventure
- 2009
(Show Context)
Citation Context ...s becoming more and more popular for applications. However, constantly probing the network is not suitable. To make measurements more scalable, the notion of clustering has emerged. We demonstrate in =-=[2]-=- that clustering can limit the measurement overhead in such a context without loosing too much accuracy. The paper shows that measurement reduction can be observed when vantage points collaborate and ... |

1 |
Triangle inequality violation avoidance in internet coordinate systems
- Liao, Leduc
- 2009
(Show Context)
Citation Context ...m is required. This significantly reduces the overhead of active probing and largely improves the efficiency of the network. We have used Machine Learning techniques to improve the accuracy of an ICS =-=[3, 4, 5]-=-. We have derived automatically a criterion that can be used by nodes to better select their neighbours in the ICS and thereby reduce the impact of Triangular Inequality Violations (TIVs), which are d... |

1 | Long horizon end-to-end dealy forecasts: a multi-step-ahead hybrid approach - Bui, Zhu, et al. - 2007 |

1 | Nonlinear modeling of the internet delay structure
- Wang, Chen
- 2008
(Show Context)
Citation Context ...of Technical Objective 2 Page 25posed in order to mitigate the impact of the shortcuts on the quality of the embedding. One of these approaches simply consists in applying a non linear transformation=-=[23]-=- like y = x 1/n (where n is a parameter) to the delays before trying to estimate them. We have observed that n = 1.5 is the value of the parameter that gives the better results according to the estima... |

1 |
Heuristics for diverse routing in wavelength-routed networks with shared risk link groups
- Pan, Xiao
(Show Context)
Citation Context ...both links share the same risk). Therefore, any form of cable cut for that particular physical link may manifest two simultaneous link failures and eventually the two links are part of the same SRLG (=-=[34]-=-, [35], [36]). Identification of SRLGs in the network topology reduces the failure recovery time and therefore total amount of packet losses ([34], [35], [36]). This is because, otherwise, each link f... |

1 |
Bundling in Optical Networks
- Link
- 2000
(Show Context)
Citation Context ...mportant aspect of network protection and restoration. There exist several methods for SRLG detection and identification as well as association of SRLGs (e.g. shared risk link grouping technique [37] =-=[38]-=-). However, the methods investigated so far are dependent on underlying physical network, e.g., optical network and requires cross layer information transfer among the network nodes. The process event... |

1 |
rate fairness: dismantling a religion
- Flow
- 2007
(Show Context)
Citation Context ...profile-based accountability use case within the context of congestion and fairness. In this context, profile-based accountability focuses on a local approach to subscriber accountability. Similar to =-=[41]-=- and [43], the proposed approach goes beyond individual flows by tracing subscriber traffic rates and by characterizing them into account the local congestion the network incurs. Different from [41] a... |

1 |
Freedom to use the internet resource pool
- Jacquet, Briscoe, et al.
- 2008
(Show Context)
Citation Context ...ased accountability use case within the context of congestion and fairness. In this context, profile-based accountability focuses on a local approach to subscriber accountability. Similar to [41] and =-=[43]-=-, the proposed approach goes beyond individual flows by tracing subscriber traffic rates and by characterizing them into account the local congestion the network incurs. Different from [41] and [43] i... |