Abstract In recent years, cluster analysis has played an increasingly important role in statistical pattern recognition. Hoeltzel and Chieng have shown an example on cognitive selection of nonlinear programming algorithms in a mechanical design expert system. In this paper, an improved dynamic clustering of 3000 samples came from a comparative performance evaluation of six typical nonlinear programming softwares with randomly generated test problems has been made. Explanations resulting from the cluster analysis have been used to build rules to form the knowledge base of an optimization expert system.