Deep Learning for Radar Waveform Design: Retrospectives and the Road Ahead

Adaptive transmission denotes the ability of the radar system to alter its transmit waveform in response to the surrounding environment. In the open literature, the waveform design process boils down to an optimization problem that maximizes the radar performance in terms of signal to interference plus noise ratio (SINR), desired transmit-receive beampattern and ambiguity function shaping etc., subject to a set of constraints that encapsulate necessary and desired characteristics of the waveform. Incorporating these constraints invariably leads to hard non-convex problems and remains a longstanding open challenge. Recent work has seen the advent of machine learning, specifically deep learning methods for the constrained radar waveform design problem, in addition to the already mature body of numerical optimization algorithms. This paper provides a review of key learning based approaches for waveform design based on canonical deep regression architectures including fully connected layer and residual network, compares and contrasts their merits and drawbacks. We contend that while deep learning methods have significant potential for improving computational speed and performance, the issues of explainability and generalizability (training robustness) must be rigorously addressed to enable reliable, practical learning-based techniques that are deployable.