Professor John R Buck at ECE Spring 2024 Colloquium

Universal Adaptive Beamforming: Machine Learning meets Array Processing

Practical adaptive beamformers operating in dynamic environments regularize the sample covariance matrix. Choosing the best regularization parameter for sequential online processing remains a difficult open challenge. The universal adaptive beamformer (UABF) combines online learning techniques with array processing to create a new beamforming framework. The UABF's array weights are a weighted average across the array weight vectors of a set competing beamformers based on their recent performance. Each competing beamformer implements a different regularization parameter choice. 

Remarkably, the performance gap between the UABF and the best of the competing beamformers chosen with 20/20 hindsight vanishes asymptotically for every individual sequence of observed data. Simulations and a data set of opportunity from the Philippine Sea experiment demonstrate the effectiveness of the UABF. [Work supported by ONR 321US]