Prediction and control in a just-in-time environment using neural networks

Abstract

The success of the Japanese just-in-time (JIT) with kanban inventory control technique has caused many manufacturing firms world-wide to implement similar systems in an attempt to remain competitive. Predicting and controlling the number of kanbans in an unstable environment is a complex decision involving many stochastic factors. This research investigates using neural computing (neural networks) to identify endogenous factors (shop conditions) and exogenous factors (product demand and supplier schedules) that are correlated with kanban system performance and to predict the optimal number of kanbans based on the "dynamic" interaction (changing over time) of these factors inherent in many production environments. The purpose of the research is to test the interpolative ability of a neLiral network to synthesize a multidimensional response surface from sample values and to perform factor screening on the inputs. First, a JIT shop simulator capable of utilizing different factor levels is used to generate data on shop performance for different kanban levels for 560 dynamic shop scenarios. Each combination of shop factor levels, along with the corresponding optimal number of kanbans, is saved in a data file. The data is randomly split into 2 files of equal size. The first file is used as training data for a neural network. The neural network "learns" the relationship between the shop factors and the correct number of kanbans needed from the training data. After the training phase, the neural network is tested on its "associative" ability to determine how well it predicts the correct number of kanbans for the shop scenarios in the second file (data it has never seen). Results are given for different network paradigms to determine the best paradigm for predicting the number of kanbans in a dynamic JIT shop. The neural network is also used as a tool for factor screening. Each factor is analyzed to determine its relative importance in kanban prediction. Statistical tests are used to gauge the importance of the dynamic information as well as to examine the relevance of various factor groupings. The results have practical implications for firms that have adopted, or are considering, the JIT technique.