A study of formal method for generating structural high leverage decision function in system dynamics models

• Main Authors: Chia-Ping Chen, 陳加屏
• Other Authors: Showing H. Young
• Format: Others
• Language: zh-TW
• Published: 1999
• Online Access: http://ndltd.ncl.edu.tw/handle/17895631351110273244

博士 === 國立中山大學 === 企業管理學系 === 87 === Designing high leverage decision function is a very crucial and challenging step in system dynamics approach. However, very limited formal methods were developed in this area. Literature showed that these methods which can obtain structural decision function are not easy to use and not suitable for nonlinear models, but these methods which are easy to use and suitable for nonlinear models can’t obtain structural decision function. So the objective of the study is to develop an easy and formal method for generating structural high leverage decision function in system dynamics models. The idea of the method came from our experimental studies of microworlds. In which we observed that if the subjects repeatedly play a microworld by trial and error, they could often implicitly learn how to control the microworld even when they did not know the underlying structure. This kind of cognitive behavior is useful for controlling system dynamics models. So we imitated it to construct a working hypothesis. And then we followed the working hypothesis to direct the development of our method. In short, this kind of cognitive behavior has two major activities: selecting information and organizing information. We adopt the genetic algorithm as the mechanism for selecting information, and adopt the back-propagation algorithm as the mechanism for organizing information. Actually, there are two stages in the operational process. The first stage is to obtain the open loop solution of the system by the Powell optimal algorithm combining Fourier series. The open loop solution is the optimal trajectory of decision point. The second stage is to obtain the closed loop solution, that is structural decision function, based on the open loop solution by GNN (Genetic Neural Network software, combining genetic algorithm and artificial neural network.) We applied the method to two different types of model: the growing oriented model--Forrester’s market growth model, and the stable oriented model--Forrester’s customer-producer-employment model. The results of two experiments showed that the performance of the decision functions obtained by our method both are better than that of Forrester’s decision functions. Then we compared our method with Macedo’s formal method. The result also showed our method better than Macedo’s. There are two implications of management developed by the study based on our method. One is to screen decision relevant information. And the other is to reconcile the dynamic complex conflicts among the hierarchy of objectives.