Machine Learning for Optimizing Cognitive Radar Waveforms

Malcarne, Joshua

Etd

Machine Learning for Optimizing Cognitive Radar Waveforms

Öffentlich Deposited

The current push for spectrum reallocation in the sub-6 GHz frequency range to support 5G development threatens to hinder the performance of legacy systems at these frequency bandwidths, such as radar systems. This study proposes a novel machine learning model to optimize radar waveforms for coexistence with 5G signals in shared frequency bands, addressing the challenge of enabling 5G development while safeguarding radar systems. The approach involves optimizing radar waveforms for transmission through unallocated slots within the 5G signal space to mitigate interference—a method not previously explored. We evaluate two techniques: Stochastic Gradient Descent (SGD) and Deep Q-Learning (DQN). The SGD approach faced significant issues with waveform output, while the DQN method showed promise, successfully converging to optimal waveforms in initial tests. However, DQN produced lower-than-expected sidelobe-to-mainlobe ratios and rewards when applied to the 5G dataset due to problems with reward calculations and sidelobe-to-mainlobe ratio processing. Despite these challenges, the DQN approach established a foundation for developing machine learning models for spectrum sharing applications, including future development of varied window functions, exploration of alternative optimization metrics like signal-to-noise ratio (SNR), and creation of a labeled 5G dataset to improve training efficacy. In addition to spectral coexistence, these same concepts and contributions can apply to other use cases without loss of generality.

Creator