Abstract
Neural networks (NNs) are tools that are very frequently successfully applied in the modeling of various phenomena and processes. This is due to combination of characteristic for NNs wide approximation capabilities (manifesting especially in nonlinear modeling tasks) with their flexibility and high performance in fitting the model to the real data during the learning process. Taken together these features make NNs one of the best modeling tools available. However, it is a common practice to achieve success with neural network technique in a modeling of particular system while confining the research only to neural model selection, optimization of parameters and validation of the NN performance goodness. Frequently, neural models predictions are analyzed and compared with other modeling techniques or other neural systems. In this paper we provide a complementary approach to the above-mentioned scheme. We took one non-trivial modeling task as an example (i.e. prediction of biological activity of chemical compounds based on their structure and properties) and studied various types of neural networks in order to determine the optimal type of NN, which deals with modeling problem in the most efficient way. We analyzed both linear and non-linear neural networks of MLP and GRNN type. In non-linear MLP systems the linear or non-linear output layers were tested. Moreover a hybrid neural system was developed that joins results of architecture optimization of MLP and GRNN. The paper addresses also the issue of input parameters selection, optimal number of hidden neurons and data representation, especially in terms of an output results. A dozen or so thousands of neural models were developed, providing a rich dataset for assessment of neural networks usefulness. It seems that such a comparative study can be of a high value for other researchers using neural systems in modeling studies. It should allow to chose a type and size of NN used based less on arbitrary and more on rational basis. Our results provide also better understanding into the character and cause-result relationship of processes that take place in neural networks