AbstractWe devised a network with the “functional‐link net” architecture (Klassen and Pao, 1988) for QSAR for activity data given by ratings. Reference values used for training of the network were uniquely defined. In other neural networks such as the generalized delta rule net, reference values for activity ratings are usually defined by patterns using the combination of 1 and 0 so that more than one node is needed in the output layer.We, however, defined the reference by real values ranging from 0.0 to 1.0, thus, only one node sufficed in the output layer. Application of this network to QSAR of three data sets, 16 and 14 mitomycin derivatives with anticancer activity and 29 arylacryloylpiperazines with antihypertensive activity, gave good results. The comparison with other methods, such as ALS, FALS and the generalized delta rule net with back‐propagation of error, showed that our network is not only characterized by good recognition, but also by high prediction