An Indicator Selection Method Based on Robust Fractal Autoencoder

Evaluation indicator selection is a key step in constructing an effectiveness evaluation indicator system and directly affects the quality of subsequent evaluation results. Existing autoencoder-based indicator selection methods mainly learn indicator contributions from reconstruction errors. In effectiveness evaluation scenarios, however, they still suffer from relatively high redundancy in selected indicators, insufficient stability of selection results, and limited adaptation to the construction requirements of evaluation indicator systems. To address these issues, this paper proposes a robust fractal autoencoder-based indicator selection method adapted to evaluation indicator systems. Building on the global reconstruction, local reconstruction, weight exponentiation, and dynamic-window mechanisms of the Robust Fractal Autoencoder (RFAE), the proposed method introduces indicator-selection-oriented constraints. These constraints regulate indicator selection weights from three perspectives: indicator redundancy, capability-dimension coverage, and sample discriminability. As a result, the selected indicators better satisfy the construction requirements of a refined, stable, reasonably covered, and interpretable effectiveness evaluation indicator system. Experiments based on multi-agent adversarial simulation data show that the proposed method can be used to construct a core indicator subset with low redundancy, reasonable coverage, and favorable sample discriminability, thereby providing a feasible approach for the automated construction of effectiveness evaluation indicator systems for complex systems.