Utilizing Machine Learning for Planning Unsupervised Future Cellular Networks
Reducing the cost of deploying cellular networks is of great interest to all concerned parties, including service providers and users. Such cost can be associated with operation (OPEX) or capital (CAPEX). Towards achieving this goal, the wireless industry is moving towards zero-touch cellular networks, i.e., zero human intervention. The need for having unsupervised (i.e., automated) cellular networks is aligned with the vision of having a dynamic cellular architecture, enabled by the use of mobile equipment (e.g., unmanned aerial vehicle base stations), which gives the architecture flexibility to adapt quickly and frequently to service demands. To this end, the concept of self-organizing network (SON) has been established and added to the list of 5G/6G key enabling technologies that aim at automating the processes of planning, configuration, management, and healing cellular networks. Among these processes, radio access network (RAN) planning has received special attention, since it decides on the required radio resources and the equipment to deploy, which directly affects CAPEX. Motivated by the above, we present in this talk a framework that aims at developing an unsupervised planning process that provides the essential planning parameters of cellular networks, including the minimum number of required base stations (BSs), their positions, coverage, and antenna radiation patterns, while taking into consideration the inter-cell interference and satisfying capacity, coverage and transmit power constraints. We make use of the statistical machine learning (SML) theory to solve the problem at hand. The core idea of SML is that the planning parameters are treated as random variables. The parameters that maximize the corresponding joint probability distribution, conditioned on observation of users’ positions, are learned or inferred using Gibbs sampling theory and Bayes theory. The inference process involves linking the observations and the planning parameters through a probabilistic model (i.e., a problem formulation) which yields a Dirichlet process. Through several numerical examples, we show that the performance of the proposed framework is superior to two existing main planning approaches, including the k-mean based approach. We also demonstrate how our approach can leverage existing cellular infrastructures into the new design.