This paper introduces a multi-level search framework for input-output configuration optimization with embedded feature selection. Building on the MICS-EFS algorithm, a hierarchical coarse-to-fine strategy is proposed to reduce the computational complexity of exploring high-dimensional configuration spaces. The method progressively refines candidate solutions across resolution levels, enabling efficient identification of informative feature subsets. The approach is evaluated on multiple benchmark datasets, including image and sensor-based data. Results show that the proposed method outperforms a shallow CNN baseline while achieving performance comparable to the original MICS-EFS algorithm. This is accomplished with significantly reduced computational cost, using only 4-8% of the input features and reducing runtime to approximately 40% of the baseline.