Cellular Network Optimization by Deep Reinforcement Learning and AI-Enhanced Ray Tracing.

In this paper we study the use of deep reinforcement learning that is supported by a ray tracer, on top of a detailed 3D model of the geospatial environment, for optimization of antenna tilts in cellular networks. We propose two novel mechanisms---geospatial importance sampling and multi-path coefficient---to efficiently pass geospatial information to the reinforcement learning model. We show that this approach can be used for fast and scalable optimization of tilt levels of cellular antennas. We present an experimental evaluation that compares the use of reinforcement learning to greedy search, simulated annealing and Bayesian optimization. Our study shows that reinforcement learning is effective and can cope with optimization problems that are at a greater scale than the settings the other algorithms can cope with.