COMPUTATIONAL NEURAL NETWORKS AS MODEL-FREE MAPPING DEVICES

Computational neural networks (CNNs) represent a set of computational paradigms with the power to address a wide range of problems from pattern recognition to system identification. The present work focuses on a number of issues that arise in essentially all CNN applications and are especially important in chemical applications of quantitative structure-activity and structure-property relationships (QSAR and QSPR, respectively). Issues related to network optimization, data representation, error analysis, and generalization are identified, and their significance to CNN applications is described. A three-dimensional response surface designed to model many problems associated with QSAR and QSPR predictions is described, and the results of extensive CNN experiments based on this surface explicitly address a number of the issues. Special emphasis is placed on the critical issues of small data sets and noisy data that plague many chemical applications of neural nets.