Machine Learning in Networking: Network Adaptation and Performance Analysis

对于网络运营商来说，5G和物联网都提供了巨大的潜力，也带来了巨大的挑战。由于预计会有大量联网设备，并且这些设备的可用带宽会有数量级的增长，网络运营商必须妥善规划其网络，以处理未来不可预见和不可预测的应用所产生的海量流量。在先进的处理能力、复杂的算法和丰富的数据的推动下，人工智能和机器学习已经走到了研究和现实问题解决应用的前沿。本研究旨在探索将机器学习技术用于流量分类和预测，以辅助网络规划和运营。 谢里登团队将与罗杰斯通信公司合作，开展以下关键研究活动。 1.分析原始4G流量数据，开发必要的工具将数据格式化和映射为合适的格式，并进行必要的预处理，为特征提取准备数据。 2.识别已知和期望的特征，探索新的和潜在的特征，创建派生特征，并开发一组特征，由Rogers审查，用于最终的模型构建。 3.构建了两个最大似然模型--一个用于流量分类，一个用于流量预测。迭代地，使用测试数据集和所选特征集来训练和调整模型。 4.为每个模型开发一组测试场景，并使用测试数据集对模型进行有条不紊的测试。使用特定于算法的标准性能评估矩阵和Rogers所需的操作性能阈值来分析模型的性能。

For network operators both 5G and IoT offer huge potential as well as immense challenge. With expected numerous network-connected devices, and orders of magnitude increased available bandwidth to these devices, network operators must plan their network properly to handle the deluge of traffic generated by unforeseen and unpredicted applications of future. Facilitated by advanced processing power, sophisticated algorithms and abundance data, artificial intelligence and machine learning has come to forefront of research and real-life problem-solving applications. This research aims to explore the use machine learning techniques for traffic classification and prediction to assist in network planning and operation. The Sheridan team, in collaboration with Rogers Communication, will carry out the following key research activities. 1. Analyze raw 4G traffic data, develop necessary tools for formatting and mapping data into suitable format, and carry out required preprocessing to prepare data for feature extraction. 2. Identify the known and desired features, explore for new and potential features, create derived features, and develop a set of features, vetted by Rogers, for final model building. 3. Construct a pair of ML models- one for traffic classification and one for traffic prediction. Iteratively, train and tune the models with the test dataset and selected feature set. 4. Develop a set of test scenarios for each model and perform methodical testing of the models with the test dataset. Analyze performance of the models with algorithm-specific standard performance evaluation matrices and required operational performance thresholds of Rogers.