A Machine Learning Enhanced Scheme for Intelligent Network Management

The versatile networking services bring about huge influence on daily living styles while the amount and diversity of services cause high complexity of network systems. The network scale and complexity grow with the increasing infrastructure apparatuses, networking function, networking slices, and underlying architecture evolution. The conventional way is manual administration to maintain the large and complex platform, which makes effective and insightful management troublesome. A feasible and promising scheme is to extract insightful information from largely produced network data. The goal of this thesis is to use learning-based algorithms inspired by machine learning communities to discover valuable knowledge from substantial network data, which directly promotes intelligent management and maintenance. In the thesis, the management and maintenance focus on two schemes: network anomalies detection and root causes localization; critical traffic resource control and optimization. Firstly, the abundant network data wrap up informative messages but its heterogeneity and perplexity make diagnosis challenging. For unstructured logs, abstract and formatted log templates are extracted to regulate log records. An in-depth analysis framework based on heterogeneous data is proposed in order to detect the occurrence of faults and anomalies. It employs representation learning methods to map unstructured data into numerical features, and fuses the extracted feature for network anomaly and fault detection. The representation learning makes use of word2vec-based embedding technologies for semantic expression. Next, the fault and anomaly detection solely unveils the occurrence of events while failing to figure out the root causes for useful administration so that the fault localization opens a gate to narrow down the source of systematic anomalies. The extracted features are formed as the anomaly degree coupled with an importance ranking method to highlight the locations of anomalies in network systems. Two types of ranking modes are instantiated by PageRank and operation errors for jointly highlighting latent issue of locations. Besides the fault and anomaly detection, network traffic engineering deals with network communication and computation resource to optimize data traffic transferring efficiency. Especially when network traffic are constrained with communication conditions, a pro-active path planning scheme is helpful for efficient traffic controlling actions. Then a learning-based traffic planning algorithm is proposed based on sequence-to-sequence model to discover hidden reasonable paths from abundant traffic history data over the Software Defined Network architecture. Finally, traffic engineering merely based on empirical data is likely to result in stale and sub-optimal solutions, even ending up with worse situations. A resilient mechanism is required to adapt network flows based on context into a dynamic environment. Thus, a reinforcement learning-based scheme is put forward for dynamic data forwarding considering network resource status, which explicitly presents a promising performance improvement. In the end, the proposed anomaly processing framework strengthens the analysis and diagnosis for network system administrators through synthesized fault detection and root cause localization. The learning-based traffic engineering stimulates networking flow management via experienced data and further shows a promising direction of flexible traffic adjustment for ever-changing environments