Applications of machine learning in cellular base station supervision

Abstract

Existing research on cellular network optimization using machine learning can generally be divided into two different methodologies: Offline learning and online learning. Offline learning trains a machine learning agent off the network using simulations or data gathered from real base stations to determine the optimal set of parameters. The resulting parameters are then applied to the network, achieving performance greater than that of common heuristic approaches. Online learning methods on the other hand train the agent on the network itself while the network is operational. The agent will then over time find an optimal set of parameters specific to that network to ensure high performance.

This thesis proposes that there exists an under-researched hybrid approach that makes use of both approaches. In it, an agent is trained offline in how to search the parameter space of the network to discover the configuration that results in the highest performance automatically. By doing this, the agent is capable of finding the global maximum of performance in the parameter space faster than online methods, making it more responsive and capable of handling changing environments.

In this thesis, this hybrid method is further explored and tested. Neuroevolution is used to create an agent to navigate a simplified model of a cellular network parameter space to optimize the throughput of the network. The results highlight the need for a specialized approach that can operate despite the lack of information about the parameter space while the system is online.