Application of global-extreme-learning to law-discovery neural networks

The problem of improving efficiency of training special-type neural networks (SNN) used to create symbolic description of rules governing a set of empirical data is considered. Values of the description parameters are determined by the network training. Difficulties with the SNN learning appear mainly due to a great number of local minima encountered in this process. The learning methods we applied so far were based on a modified version the Back Propagation algorithm called BP-CM-BFGS. It turned out, however, that this approach is not always effective, especially when the number of input variables of the SNN increases. In this paper, we propose to use as training technique an evaluation algorithm called Differential Evolution (DE). To illustrate effectiveness of this technique we present results of learning a reciprocal-function-based SNN [15] implementing a fifth order polynomial.