Efficiency evaluation serves as a critical tool in modern management systems, providing an objective measure of the resource conversion efficiency, output quality, and value realization level of organizations, projects, teams, and individuals through standardized indicator design and normalized data analysis. However, with the rapid development of the economy and technology, the volume of data requiring evaluation has increased dramatically. Under these circumstances, it is essential to analyze and process massive amounts of data indicators to achieve accurate efficiency assessment. This paper studies methods for simplifying efficiency evaluation indicator systems. By employing an autoencoder neural network approach, multidimensional indicator data—which is often high-dimensional, redundant, and contaminated with noise—is reduced in dimensionality. This allows a smaller set of characteristic indicators to replace the original evaluation system, thereby reducing computational complexity and extracting essential features. The extracted key features enable accurate efficiency evaluation, forming an effective method for simplifying efficiency evaluation indicator systems. This approach breaks the limitations of experience-driven decision-making and provides a scientific basis for multi-agent decision-making.