[[alternative]]A Genetic Algorithm Approach for the Tool Portfolio Planning

[[abstract]] 半導體產業機台設備非常昂貴，且製程加工的過程也十分繁複，無法以傳統單一生產線的方式進行，所以機台組態配置對於半導體產業而言，是非常重要且複雜的工作。本論文使用混合式演算法，以線性規劃搭配基因演算法，求取機台組態配置的較佳解。 在本研究中，我們提出二種調整基因演算法之方法，分別為初始染色體種子法與持續染色體種子法，初始染色體種子法是在初始母體中加入種子染色體，持續染色體種子法則是在各個子代中加入種子染色體。我們利用基因演算法易落入區域最佳解之特性，加入一個種子染色體，而此種子染色體則是一組限制機台產能上限的最佳解，以此最佳解誘導在基因演算法求解時往最佳解方向搜尋。 本研究的實驗設計為，比較傳統基因演算法與二種調整後之基因演算法，透過實驗結果證實，初始染色體種子法與持續染色體種子法，在五組數據中平均最佳解皆高於傳統基因演算法，故可證明本研究提出之方法具實用價值。[[abstract]] The object of this study is to develop a revised hybrid genetic algorithm for the semiconductor tool portfolio allocation problem. Due to the complicated processes in the semiconductor industry, it is not easy to check and allocate the required tool capacity to each related processes. This may result in failure of capacity check and inefficient allocation of tool capacity. However, as the tool capacity ceiling varies as the tool is used in single process or been used across processes, this problem requires a complex mixed integers programming to solve it. So it is wonder if a hybrid GA along with LP method could be used in the problem and this study proposed a revised GA (Seeded GA) to gain the better solution. And a seed chromosome was proposed in this study. The gene in this GA method represents the status whether a tool is used in a certain process. And the seed chromosome is derived from the optimal solution by performing linear programming to the simplified tool portfolio allocation problem. The simplified problem was designed as tool capacity is fixed as 0.9, which is 90% of the full capacity. The seed chromosome was arranged in the first parent pool. The fitness value is determined by the linear programming of the problem formed by the chromosome. It is hoped that the seed could lead the evolution process to find some better solution. Two variations were designed: the seed was arranged in the first population pool only or stays in each iteration chromosome pool. Five MPS were used in the experiment and the results showed that Seeded GA could find better solution in very few iteration. However, there is no much difference among these two variations