Optimizing the fuzzy-nets training scheme using the Taguchi Parameter Design
Nan-Hui Lin · 1996
Neural networks and fuzzy logic have attracted much attention and interest worldwide for several decades. Both technologies have been widely applied in many domains including industrial, business, and academic fields. The most recent approach, called fuzzy-nets, combines the learning ability of neural networks and the simplicity of fuzzy logic to deal powerfully with complex control systems. How can fuzzy-nets technology perform with greater accuracy? It is unclear how the possible factors affect the performance of fuzzy-nets. The purpose of this study was to identify the optimal factor-level combination of a simulation model using the backing up of a truck. The four factors considered, each at three levels, were: (a) the number of training sets; (b) the number of fuzzy regions; (c) the membership functions; and (d) the fuzzy reasoning methods. Taguchi Parameter Design with L₉ (3⁴) orthogonal array was employed to reduce the number of treatment runs. Step-by-step sequences were followed for designing, analyzing, and interpreting the experiment in this research. All raw data were analyzed statistically by analysis of variance (ANOVA) in order to determine the significant factors. Then SIN ratios optimized the mean and minimized the deviation simultaneously. Following construction of the response tables and executing confirmation runs, the optimal level combination was identified based on fewer data sets, three fuzzy regions, a trapezoidal membership function, and a square product reasoning. The performances of the proposed fuzzy-nets scheme were represented by the average errors between a truck and loading dock, 0.178 units and 0.204 degrees. These small values indicate that the optimal factor-level combination was quite reasonable and accurate.