Towards self-driving radios: Physical-layer control using deep reinforcement learning

Abstract

Modern radios, such as 5G New Radio, feature a large set of physical-layer control knobs in order to support an increasing number of communication scenarios spanning multiple use cases, device categories and wireless environments. The challenge however is that each scenario requires a different control algorithm to optimally determine how these knobs are adapted to the varying operating conditions. The traditional approach of manually designing different algorithms for different scenarios is increasingly becoming not just difficult to repeat but also suboptimal for new scenarios that previous-generation radios were not designed for. In this paper, we ask: can we make a radio automatically learn the optimal physical-layer control algorithm for any scenario given only high-level design specifications for the scenario, i.e., can we design a self-driving radio? We describe how recent advances in deep reinforcement learning can be applied to train a self-driving radio for several illustrative scenarios, and show that such a learning-based approach not only is easily repeatable but also performs closer to optimal than the current state of the art.