Machine learning for wireless networking has the potential to provide a breakthrough in network management and optimization.This article, details the basic workflow of how the basic workflow model can be applied in the network context.
The main phases involve problem formulation, monitoring and data collection, data processing and analysis, model construction, model tuning, model deployment and inference.
Although these phases have prominent correlations and inter-dependencies, this workflow provides a high-level distinction between the main components required to apply machine learning in data networks.
Problem formulation: In this phase, the network phenomenon under scope is logically analyzed to identify the key parameters that drive its behavior. Among others, this phase requires sufficient abstraction and meticulous formulation of the network problem under scope in order to infer on the amount of data to collect, the learning models that are more relevant to the problem under scope as well as the machine learning approaches to be employed (e.g. classification, clustering, etc). The degree of regression is also an important parameter to decide on at this stage, based on the importance of online data as compared to the exploitation of previous historic data and actions taken.
Monitoring and data collection: During this phase it is critical to identify establish the channels required to collect the vast volume of network representation data (e.g. traffic patterns, load distribution, performance logs, failure logs, security analytics, etc) that are transferred throughout both the physical and virtual network infrastructure.
Data processing and analysis: At this stage, raw data should be filtered, classified and utilized so as to extract features on the dynamics, evolvement and behavior of the network phenomenon under scope. For example, the set of input data that play an important role on the video playback experience of the users should be processed here so as to provide efficient means in the decision phase to adapt the numerous video playback parameters (e.g. coding rate, video format, etc).
Model construction: In this phase, data analytics areutilized so as to construct the machine learning model, balance between the available online and offline data, adjust the regression rates and manage the output provided by the model. The key target is to provide ready-to-utilize output in terms of potential actions to implement or classification of the network state in light of the next module that will utilize the deployed prediction models.
Model tuning: Among the targets of this phase, is the fine-tuning of the model by comparing network performance with the model prediction of previous stages. Optimization of the model and adjustment of its parameter will take place at this stage, meaning that the sampling rates, richness of data, amount of information should also be adjusted accordingly. .
Model deployment and inference: This phase entails the actual implementation of the learning model in real-life conditions, taking into account the accuracy level required, the trade-offs between accuracy and overhead paid, the relative gains attained with and without the employment of the proposed learning model, etc.