Prediction in real-time control using adaptive networks with on-line learning

Abstract

Adaptive systems are useful in many process control applications. Especially neurofuzzy systems are of interest because they may be applicable in safety-critical domains. But to cope with large input numbers, it is necessary to split such systems into a network. Such an approach, the NeuroFuzzy Network (NFN), is outlined. Its use is demonstrated by modeling a biological reactor in order to use a one-step prediction for correcting destroyed measurement values. The training of the NFN is done on-line by exploiting the power of a multiprocessor system. Investigations show the improvements and limitations of parallel processing for on-line learning in adaptive networks.<>