Abstract: We propose a learning-based DOA framework that selects subarray elements dynamically to address array distortion and path-dependent loss. A lightweight classifier is trained to activate ...