Online Meta-Learning for Scene-Diverse Waveform-Agile Radar Target Tracking

Abstract: A fundamental problem for waveform-agile radar systems is that the true environment is unknown, and transmission policies which perform well for a particular tracking instance may be sub-optimal for another. Additionally, there is a limited time window for each target track, and the radar must learn an effective strategy from a sequence of measurements in a timely manner. This paper studies a Bayesian meta-learning model for radar waveform selection which seeks to learn an inductive bias to quickly optimize tracking performance across a class of radar scenes. We cast the waveform selection problem in the framework of sequential Bayesian inference, and introduce a contextual bandit variant of the recently proposed meta-Thompson Sampling algorithm, which learns an inductive bias in the form of a prior distribution. Each track is treated as an instance of a contextual bandit learning problem, coming from a task distribution. We show that the meta-learning process results in an appreciably faster learning, resulting in significantly fewer lost tracks than a conventional learning approach equipped with an uninformative prior.