Packet-loss prediction model based on historical symbolic time-series forecasting
- Publication Type:
- Issue Date:
Rapid growth of Internet users and services has prompted researchers to contemplate smart models of supporting applications with the required Quality of Service (QoS). By prioritising Internet traffic and the core network more efficiently, QoS and Traffic Engineering (TE) functions can address performance issues related to emerging Internet applications. Consequently, software agents are expected to become key tools for the development of future software in distributed telecommunication environments. A major problem with the current routing mechanisms is that they generate routing tables that do not reflect the real-time state of the network and ignore factors like local congestion. The uncertainty in making routing decisions may be reduced by using information extracted from the knowledge base for packet transmissions. Many parameters have an impact on routing decision-making such as link transmission rate, data throughput, number of hops between two communicating peer end nodes, and time of day. There are also other certain performance parameters like delay, jitter and packet-loss, which are decision factors for online QoS traffic routing. The work of this thesis addresses the issue of defining a Data Mining (DM) model for packet switching in the communications network. In particular, the focus is on decision-making for smart routing management, which is based on the knowledge provided by DM informed agents. The main idea behind this work and related research projects is that time-series of network performance parameters, with periodical patterns, can be used as anomaly and failure detectors in the network. This project finds frequent patterns on delay and jitter time-series, which are useful in real-time packet-loss predictions. The thesis proposes two models for approximation of delay and jitter time-series, and prediction of packet-loss time-series – namely the Historical Symbolic Delay Approximation Model (HDAX) and the Data Mining Model for Smart Routing in Communications Networks (NARGES). The models are evaluated using two kinds of datasets. The datasets for the experiments are generated using: (i) the Distributed Internet Traffic Generator (D-ITG) and (ii) the OPNET Modeller (OPNET) datasets. HDAX forecasting module approximates current delay and jitter values based on the previous values and trends of the corresponding delay and jitter time-series. The prediction module, a Multilayer Perceptron (MLP), within the NARGES model uses the inputs obtained from HDAX. That is, the HDAX forecasted delay and jitter values are used by NARGES to estimate the future packet-loss value. The contributions of this thesis are (i) a real time Data Mining (DM) model called HDAX; (ii) a hybrid DM model called NARGES; (iii) model evaluation with D-ITG datasets; and (iv) model evaluation with OPNET datasets. In terms of the model results, NARGES and HDAX are evaluated with offline heterogeneous QoS traces. The results are compared to Autoregressive Moving Average (ARMA) model. HDAX model shows better speed and accuracy compared to ARMA and its forecasts are more correlated with target values than ARMA. NARGES demonstrates better correlation with target values than ARMA and more accuracy of the results, but it is slower than ARMA.
Please use this identifier to cite or link to this item: