Response surface methodology (RSM) is an important tool for process parameter optimization, robust design and other quality improvement efforts. When the relationship between influential factors and output response is very complex, it's hard to find the real response surface by using RSM. In recent years Artificial Neural Networks (ANN) has been used in RSM. But the classical ANN does not work well under the constraints of real applications. An algorithm is proposed in this paper, which is a supplement of the classical ANN methodology. It makes networks closer to the response surface, so that training time is reduced and robustness is strengthened. The procedure of improving ANN by regressions is put forward and the comparisons among the three methods are computed graphically in three examples. Our research shows that the regression-based ANN methodology is a good supplement to the RSM and classical ANN methodology, which can yield satisfied results under condition that the scope of experiments is rigidly restricted.

During process design or process optimization, it is quite common for experimenters to find optimum operating conditions for several responses simultaneously. The traditional multiresponse surfaces optimization methods do not consider the uncertain relationship among these responses sufficiently. For this reason, the authors propose an optimization method based on evidential reasoning theory by Dempster and Shafer. By maximizing the basic probability assignment function, which indicates the degree of belief that certain operating condition is the solution of this multiresponse surfaces optimization problem, the desirable operating condition can be found.