Content uploaded by Tiago prado Oliveira
Author content
All content in this area was uploaded by Tiago prado Oliveira on Feb 22, 2018
Content may be subject to copyright.
A preview of the PDF is not available
Computer network traffic prediction: A comparison between traditional and deep learning neural networks
Abstract and Figures
This paper compares four different artificial neural network approaches for computer network traffic forecast, such as: 1) multilayer perceptron (MLP) using the backpropagation as training algorithm; 2) MLP with resilient backpropagation (Rprop); (3) recurrent neural network (RNN); 4) deep learning stacked autoencoder (SAE). The computer network traffic is sampled from the traffic of the network devices that are connected to the internet. It is shown herein how a simpler neural network model, such as the RNN and MLP, can work even better than a more complex model, such as the SAE. Internet traffic prediction is an important task for many applications, such as adaptive applications, congestion control, admission control, anomaly detection and bandwidth allocation. In addition, efficient methods of resource management, such as the bandwidth, can be used to gain performance and reduce costs, improving the quality of service (QoS). The popularity of the newest deep learning methods have been increasing in several areas, but there is a lack of studies concerning time series prediction, such as internet traffic.
Figures - uploaded by Tiago prado Oliveira
Author content
All figure content in this area was uploaded by Tiago prado Oliveira
Content may be subject to copyright.
... Regarding the particular techniques exploited for traffic prediction, Tab. 1 highlights a rising utilization of DL models, also in a multitask configuration [9,11,22]. Particularly, related works mostly employ CNN of different dimensions [9,15], LSTM [9,15,17,18,19,21,23], GRU [9,17,23], SAE [13], GNN [24], and hybrid architectures obtained via their combinations [9,15,20]. Fewer works leverage Markov models (e.g., MC, HMM, and MMG) [11,14,16,22], traditional ML models (e.g., LR, SVR, k-NNR, or RFR) [9,16,18,22], and statistical techniques (e.g., ARIMA or FARIMA) [12,15,17,18], usually as performance baselines to evaluate DL models, with the latter commonly showing better prediction performance. ...
... The works performing coarse-grained evaluation (CGeval) forecast various traffic aggregates, such as bit rates [23], packet distributions [17], and traffic volumes [12,13,17,18,19,20,24] at different time resolutions, ranging from less than one second to few seconds, minutes, hours, and even days. Among the latter, some works take into account also the geographical or topological distribution of sources and destinations, rather than leveraging the (sole) temporal information, via traffic matrices [18] or considering the geographic distribution of data volumes (e.g., aggregated data calls) as observed at base stations [15,20,24]. ...
Significant transformations in lifestyle have reshaped the Internet landscape, resulting in notable shifts in both the magnitude of Internet traffic and the diversity of apps utilized. The increased adoption of communication-and-collaboration apps, also fueled by lockdowns in the COVID pandemic years, has heavily impacted the management of network infrastructures and their traffic. A notable characteristic of these apps is their multi-activity nature, e.g., they can be used for chat and (interactive) audio/video in the same usage session: predicting and managing the traffic they generate is an important but especially challenging task. In this study, we focus on real data from four popular apps belonging to the aforementioned category: Skype, Teams, Webex, and Zoom. First, we collect traffic data from these apps, reliably label it with both the app and the specific user activity and analyze it from the perspective of traffic prediction. Second, we design data-driven models to predict this traffic at the finest granularity (i.e. at packet level) employing four advanced multitask deep learning architectures and investigating three different training strategies. The trade-off between performance and complexity is explored as well. We publish the dataset and release our code as open source to foster the replicability of our analysis. Third, we leverage the packet-level prediction approach to perform aggregate prediction at different timescales. Fourth, our study pioneers the trustworthiness analysis of these predictors via the application of eXplainable Artificial Intelligence to (a) interpret their forecasting results and (b) evaluate their reliability, highlighting the relative importance of different parts of observed traffic and thus offering insights for future analyses and applications. The insights gained from the analysis provided with this work have implications for various network management tasks, including monitoring, planning, resource allocation, and enforcing security policies.
... Feed-forward MLPs are devoid of cycles and loops. Traditional MLPs typically consist of no more than three layers; beyond three layers, the MLP transforms into a deep neural network [60]. The simple MLP has several advantages, including its simplicity of implementation, superior performance, and shorter training time. ...
Vibration-based fault diagnosis from rotary machinery requires prior feature extraction, feature selection, or dimensionality reduction. Feature extraction is tedious, and computationally expensive. Feature selection presents unique challenges intrinsic to the method adopted. Nonlinear dimensionality reduction may be achieved through kernel transformations , however there is often a trade-off in information to achieve this. Given the above, this study proposes a novel autoencoder (AE) pre-processing framework for vibration-based fault diagnosis in wind turbine (WT) gearboxes. In this study, AEs are used to learn the features of WT gearbox vibration data while simultaneously compressing the data, obviating the need for costly feature engineering and dimensionality reduction. The effectiveness of the proposed framework was evaluated by training genetically optimized linear discriminant analysis (LDA), multilayer perceptron (MLP), and random forest (RF) models, with the AE's latent space features. The models were evaluated using known classification metrics. The results showed that the performance of the models depends on the size of the AE's latent space. As the size of the AE's latent space increased, the quality of features extracted improved until a plateau was observed at a latent space dimension of 10. The AE pre-processed genetically optimized RF, MLP, and LDA models, designated AE-Pre-GO-RF, AE-Pre-GO-MLP, and AE-Pre-GO-LDA, were evaluated for accuracy, sensitivity, and specificity in the classification of seven (7) gearbox fault conditions. The AE-Pre-GO-RF model outperformed its counterparts, scoring 100% for all evaluated metrics, though with the longest training time (239.50 sec). Comparable results were found comparing this study with similar investigations involving traditional vibration processing techniques. More so, it was established that effective fault diagnosis of the WT gearbox can be achieved through manifold learning with AEs without expensive feature engineering. ARTICLE HISTORY
... (2) Through adversarial learning amongst agents, Enhance the capacity for detection of a tiny quantity of data from samples. [20] find the effective resource organization approaches, Bandwidth, for example, can be used to improve efficiency while cutting costs and improving quality of service (QoS). The latest deep learning algorithms are becoming increasingly prominent in the above survey, although here is a paucity of study on time sequence calculation, such as internet traffic. ...
City traffic congestion can be reduced with the help of adaptable traffic signal control system. The technique improves the efficiency of traffic operations on urban road networks by quickly adjusting the timing of signal values to account for seasonal variations and brief turns in traffic demand. This study looks into how adaptive signal control systems have evolved over time, their technical features, the state of adaptive control research today, and Control solutions for diverse traffic flows composed of linked and autonomous vehicles. This paper finally came to the conclusion that the ability of smart cities to generate vast volumes of information, Artificial Intelligence (AI) approaches that have recently been developed are of interest because they have the power to transform unstructured data into meaningful information to support decision-making (For instance, using current traffic information to adjust traffic lights based on actual traffic circumstances). It will demand a lot of processing power and is not easy to construct these AI applications. Unique computer hardware/technologies are required since some smart city applications require quick responses. In order to achieve the greatest energy savings and QoS, it focuses on the deployment of virtual machines in software-defined data centers. Review of the accuracy vs. latency trade-off for deep learning-based service decisions regarding offloading while providing the best QoS at the edge using compression techniques. During the past, computationally demanding tasks have been handled by cloud computing infrastructures. A promising computer infrastructure is already available and thanks to the new edge computing advancement, which is capable of meeting the needs of tomorrow's smart cities.
... Each autoencoder acts as a layer-building block [59], [60], [61], [62]. In other words, SAEs can be viewed as a simple form of ANNs to reproduce input data with reduced or the same dimensionality [37], [63], [64], as shown in Fig. 3. ...
Traffic control and management applications require the full realization of traffic flow data. Frequently, such data are acquired by traffic sensors with two issues: it is not practicable or even possible to place traffic sensors on every link in a network; sensors do not provide direct information about origin–destination (O–D) demand flows. Therefore, it is imperative to locate the best places to deploy traffic sensors and then augment the knowledge obtained from this link flow sample to predict the entire traffic flow of the network. This article provides a resilient deep learning (DL) architecture combined with a global sensitivity analysis tool to solve O–D estimation and sensor location problems simultaneously. The proposed DL architecture is based on the stacked sparse autoencoder (SAE) model for accurately estimating the entire O–D flows of the network using link flows, thus reversing the conventional traffic assignment problem. The SAE model extracts traffic flow characteristics and derives a meaningful relationship between traffic flow data and network topology. To train the proposed DL architecture, synthetic link flow data were created randomly from the historical demand data of the network. Finally, a global sensitivity analysis was implemented to prioritize the importance of each link in the O–D estimation step to solve the sensor location problem. Two networks of different sizes were used to validate the performance of the model. The efficiency of the proposed method for solving the combination of traffic flow estimation and sensor location problems was confirmed from a low root-mean-square error with a reduction in the number of link flows required.
In recent decades, there has been substantial population growth, leading to a higher volume of vehicles on the roadways. This has contributed to traffic congestion issues, affecting not just major metropolitan areas but also medium-sized and small cities worldwide. The management of roadway traffic is enhanced by accurate short-term traffic flow forecasts, which makes it a crucial part of intelligent transportation systems. This study utilizes Gaussian process regression (GPR) to predict the road traffic flow under heterogeneous conditions for 5 min in the future using past data. GPR model represents the relationship between data points as a probability distribution over functions, rather than a single deterministic function as in traditional linear regression. This allows GPR to capture both the mean and uncertainty of predictions. All of the comparable models were trained and tested on actual data sets that were gathered through field research. Results of the GPR model were compared with other traditional models like autoregressive moving average model, multi-layer perceptron and cascade forward backpropagation. The performance analysis was done and the GPR model was found to be quite effective followed by other traditional neural networks. Study results confirm that the GPR model can be successfully applied for short-term traffic flow prediction under heterogeneous traffic flow conditions.
In this paper, we propose a dynamic bandwidth allocation with high utilization (DBAHU) algorithm in order to utilize the unused bandwidth of a service class. DBAHU is based on a simple and feasible dynamic bandwidth allocation (SFDBA) algorithm. Like SFDBA, DBAHU uses a common available byte counter and a common down counter for multiple queues of a service class. However, to utilize the unused bandwidth of a service class, an available byte counter can be negative unlike SFDBA. Also, the unused remainder of an available byte counter of a service class is added to available byte counters of other service classes who require more bandwidth. Using simulations, we show that DBAHU is superior to SFDBA in mean delay and frame delay variance.
Traffic load in the aggregate class of DiffServ network varies dynamically. However, when the traffic load proportion of different aggregate classes is not equal to the assigned weight proportion in the scheduling algorithm, such as the WFQ algorithm, the packets in different aggregate classes will receive unfair treatment, even though these classes have the same priority. Hence, research on fairness-oriented dynamic bandwidth allocation in DiffServ becomes very important. In this paper, in order to balance the packet loss for fair bandwidth allocation, we propose a traffic load-based dynamic bandwidth allocation approach, which especially considers both the current weight proportion and queue increment proportion for calculating the new weight proportion. Experiment shows the valid of our approach.
Forecasting exchange rates is an important financial problem which has received much attention. Nowadays, neural network has become one of the effective tools in this research field. In this paper, we propose the use of a deep belief network (DBN) to tackle the exchange rate forecasting problem. A DBN is applied to predict both British Pound/US dollar and Indian rupee/US dollar exchange rates in our experiments. We use six evaluation criteria to evaluate its performance. We also compare our method to a feedforward neural network (FFNN), which is the state-of-the-art method for forecasting exchange rate with neural networks. Experiments indicate that deep belief networks (DBNs) are applicable to the prediction of foreign exchange rate, since they achieve better performance than feedforward neural networks (FFNNs).
This article presents three methods to forecast accurately the amount of traffic in TCP/IP based networks: a novel neural network ensemble approach and two important adapted time series methods (ARIMA and Holt-Winters). In order to assess their accuracy, several experiments were held using real-world data from two large Internet service providers. In addition, different time scales (5 min, 1 h and 1 day) and distinct forecasting lookaheads were analysed. The experiments with the neural ensemble achieved the best results for 5 min and hourly data, while the Holt-Winters is the best option for the daily forecasts. This research opens possibilities for the development of more efficient traffic engineering and anomaly detection tools, which will result in financial gains from better network resource management.
The Ethernet Passive Optical Network (EPON) combines both inexpensive Ethernet equipments and the large bandwidth offered by the optical fiber. Therefore, EPON has been considered as an attractive solution for the next generation broadband access network. In this paper, we present an enhanced Dynamic Bandwidth Allocation (eDBA) algorithm for EPON network. Our mechanism allocates effectively and fairly the transmission bandwidth to ONUs. Some other algorithms, such as the Dynamic Bandwidth Allocation (DBA), introduce an idle period during the time that the Optical Line Terminal (OLT) is executing the bandwidth assignment procedure. During this idle time, no data are sent by any ONU node to OLT. In order to exploit this wasted bandwidth, the proposed algorithm calculates a vector of complementary bandwidth amounts to assign to the ONUs during the idle time. Finally, we simulate a multiservice-based EPON network configured with the eDBA mechanism. Obtained results are compared to the ones obtained with two other algorithms: Dynamic Bandwidth Allocation (DBA) and the Interleaved Polling with Adaptive Cycle Time (IPACT). We show that eDBA can significantly improve the network performance in term of packet access delay and packet loss rate as compared with other algorithms.