New Trends in Neural and Fuzzy Systems

In the presentation other than MLP neural networks are discussed. It is shown that it is possible to obtain the same function with a significantly smaller number of neurons, if ACN arbitrarily connected neurons are used. In some cases, the number of neurons could be reduced by half. Importantly, these simple architectures can also be trained faster. Unfortunately, most of popular neural network learning software is not able to handle ACN networks. A new neural network learning software is presented which works not only on arbitrary network architectures (including MLP) but neural networks can be trained about 100 times faster than if EBP algorithm is used. There are, of course, several neural network architectures which need not to be trained, or training process is minimal. These architectures usually employ a larger number of neurons; there are no training problems. These special network architectures are discussed in detail and compared. In the case of fuzzy systems the smooth control surface can be obtained with fuzzy neural networks, but again the architecture and processing time of these networks can be significantly reduced with the presented new approach to neuron-fuzzy systems.