Process Distributed Fitting and Simulation Based on Niche Genetic Algorithm

[[abstract]]基於生產計畫，模擬可提供對製程產出預測的估計，以用於評估並調整原計 畫。其中，各製程生產週期時間的估計為模擬預測中最重要部分，具正確的生產 週期時間方能進行有效的模擬。然而當製程具循序串級行為，且這些行為分佈殊 異時，目前典型配適方法如MLE僅適於單一分佈，即若EM 亦僅能處理相同分 佈的混合模型，對於不同且串級製程行為分佈，實缺乏有效的分佈參數配適方法。 由於不同串級分佈參數組合可構成相近的串級行為分佈，而這些串級分佈參 數組合將可構成所謂的多目標的柏拉前緣。因此，為有效配適製程的多分佈參 數，本研究提出一基於小生境基因演算法之多目標最佳化配適程序，因製程行為 週期時間實包含資訊傳輸延遲時間、物料等待時間與實際加工時間等三部分，若 以不同基因組代表不同串級分佈參數，藉由多組動態隔離小生境的基因交配與演 化，在柏拉前緣上，初期採用扶幼保護，中期採取新老更替與劣種不活的策略， 最後運用大數法則，將可獲得製程多分佈參數的最佳化組合。 在配適結果上，若以8 組不同串級行為分佈組合為例，其配適與實際重疊率 為84%以上。利用此程序，以8 組不同實際生產週期時間分析，與所配適參數重 建後的串級行為相較，當實際生產週期時間介於2 5 天到30 天時，參數重建後 之生產週期時間分布誤差均在0 5 天以內。因此，當進一步應用此分佈配適程序 於製程模擬時，所配適而得的製程生產時間將可提供產品於製程的可能加工時 間，以提供有效的生產模擬。[[abstract]]Using a production plan production simulation assists users to predict process outputs for evaluating and adjusting the plan The simulation efficiency depends on the correct estimation of the process cycle time for effective prediction However when there are cascaded and various behaviors distributions existed in a process the conventional distribution fitting method such as MLE suited for single distribution even EM for identical mixture model Currently an effective parameter distribution fitting is needed for various cascaded distributions If the distributions from different cascaded distributions combinations are similar to the fitted distribution then these cascaded distributed parameters will form a Pareto front Therefore this research proposes a multi-parameters optimal fitting scheme based on a niche genetic algorithm Due to a process behavior including message transmission delay material waiting time and actual processing time let the different chromosomes representing different distributed parameters evolve in multiple isolated niches and adopt the immature protection in the initial stage the new superseding and the worst die in the middle stage the large number law in the final stage then the optimal combination of the process multi-parameters is derived In fitting result given 8 cascaded distributed cases the fitting over rates excess 84% Using this proposed scheme to analyze 8 different actual process times the distributed errors are less than 0 5 days when the actual process cycle times are between 2 5 to 30 days Therefore we also applied the scheme to find the feasible processing time to simulate the Petri-Nets based processes for find an effective production output