以神經網路做混合生產線產品投入順序之研究

Abstract

本論文首先以文獻探討說明豐田所使用的目標追蹤法 GCI 及 GCII非最佳 的原因，GCI 及GCII 是利用使每一生產步驟的零件使用數量與線性增加 的零件數量差距達最小的方式，來達到 TSD 最小的目標，但因為它並不 考慮此階段投入的產品，對下一步驟產品投入時，對差距產生的影響，所 以它並不是最佳的排序法。神經網路中之 Hopfield Net 可解決組合最佳 化問題，而排序亦是組合最佳化問題，所以 Hopfield Net 可用以解決混 合生產線產品投入順序的問題。但首先要導出對應排序問題的神經腱強度 及自偏壓輸入，才得以利用 Hopfield Net 的更新運作使排序結果的 TSD 達最小。在本研究中先構建兩個用以解決排序問題的神經網路，再加 以合成後得到一組神經腱強度及自偏壓輸入，使得排序問題得以利用神經 網路加以解決。另外在執行神經網路時，採用了 Gaussian machine model 以避免神經網路陷入了所謂局部極小問題，同時採用非同步轉變模 式，以克服 Tix,ix = 0 的限制。在數值實驗中觀察得到一種從未在文獻 中提出的更新模式，此更新模式可使排序效果更佳，而這也是本研究的一 小小貢獻。最後以數值實驗結果證實了 Hopfield Net 確實可用以解決混 合生產線產品投入順序問題，且排序結果較其它排序法的排序結果為佳， 因而間接的證實本論文所推導的神經腱強度 Tix,jy 及自偏壓輸入 Tix ，皆為正確的結果，而這是本論文的一主要貢獻。 In this thesis, we first discuss why two 'goal chasing' heuristics developed at Toyota is not optimum by literature survey. GCI and GCII focus on constant component usage in each stage in order to achieve the goal of minimum TSD. But these two algorithms do not consider the effect of deviation on future stage of this stage current decision; therefore, GCI and GCII is myoptic and is not optimum. Hopfield and Tank (1985) showed that some combinatorial optimization problems can be solved by using artificial neural network systems. Because the scheduling problem is also combinatorial optimization problems, we can solve the problem of scheduling mixed-model assembly lines by Hopfield Net. In this thesis, we construct two neural networks, combine them to solve the problem of scheduling mixed- model assembly lines, and thus get a set of synaptic interconnection strength and self bias input. In this thesis, we use Gaussian machine model to avoid problems of being trapped in a local optimum solution. We also adopt asynchronous transition mode for the purpose of reducing oscillatory behaviour. In the numerical experiment, we find a renewed model which can achieve much better schedule results, and this may be considered a little tiny contribution. Finally we prove that the problem of scheduling mixed-model assembly lines can be solved by Hopfield Net via the numerical experiment, and Hopfield Net gets even better results than other scheduling algorithms. Accordingly, we have come to a conclusion that the synaptic interconnection strength and self bias input developed by the author are correct. This is the major contribution of the thesis.