Training of robust neural controllers

Controllers based on neural networks and fuzzy logic are often said to be robust, i.e., to continue to exhibit satisfactory performance even in the face of changes in the system for which they have been trained or designed. Our own experience, largely but not entirely based on model-based simulation studies, suggests that this statement carries a considerable degree of truth. In particular, we have found that the performance of dynamic control structures in the form of recurrent neural networks tends to degrade gracefully as the system is altered from that used during training. We have observed, however, that the degree of robustness does not always correlate positively with the quality of control on the plant used in training. For example, two neural controllers of identical architecture and trained on the same plant to comparable error may, on an independent test, exhibit significantly different degrees of robustness. This has led us to devise a procedure, called multistream training, that makes robustness an explicit goal of training. In the preferred form of this procedure, we train a single controller simultaneously on several plants selected from the expected range, thereby countering the tendency of a serial training process to favor the plant upon which training has most recently been conducted. This paper describes the multistream method and illustrates its use and efficacy with an example