Learning-based query optimization for multi-probe approximate nearest neighbor search