整合機器學習方法於決策樹為基智慧型排程系統之研究

Abstract

本論文目的在設計一個智慧型排程系統（ISS）。其所應具備的功能為可根據不同的生產績效指標需求，產生適時控制（派工）決策，以輔助支援現場控制系統（SFCS）之排程機制所需。由於現場控制系統中存在許多生產資訊，因此如何根據不同的生產績效指標需求，選擇出適當的生產資訊（屬性），以建構智慧型排程系統的知識庫，是一項關鍵的議題。目前在此領域的研究通常是以決策樹（DT）學習方法最為廣泛使用且為可行，然而傳統的決策樹學習方法在建構知識庫時並不重視問題領域中所存在之無關或重複的屬性，而此一問題在現場控制系統通常是顯而易見的。為解決此一問題，本論文提出兩個不同方法，利用選擇現場中基本的屬性以構建在現場控制系統中之智慧型排程系統知識庫。第一種方法為根據類神經網路（ANN）權重來衡量重要生產屬性，然後使用決策樹學習的演算法來強化知識學習機制；此一結合類神經網路及決策樹的方法稱之為ANN/DT為基之智慧型排程系統。第二種方法為遺傳演算法（GA）與決策樹學習方法之整合，從可能收集的現場資訊中以遺傳演化的方式來產生最佳組合之現場屬性子集合，以建構智慧型排程系統知識學習機制；此一結合遺傳演算法及決策樹學習的方法稱之為GA/DT為基之智慧型排程系統。實驗結果顯示，使用上述兩種不同的方法所建構之決策樹為基智慧型排程系統，在各種生產績效指標的知識歸納能力（控制決策的預測精確率）均優於傳統未經由特徵（屬性）選擇的決策樹為基智慧型排程系統。更進一步模擬實驗顯示，以本論文所提出的兩種方法所建構之決策樹為基智慧型排程系統，相對於傳統的決策樹為基智慧型排程系統及啟發式單一派工法則，長期而言在各種生產績效指標，對生產系統績效提升具有更顯著的成效。

This study develops an intelligent scheduling system (ISS) to support a shop floor control system (SFCS) to make on-line decisions, robust to various production requirements. Selecting essential system attributes (or features) based on various production requirements to construct ISS knowledge bases is a critical issue because of the existence of much shop floor information in a SFCS. However, classical decision tree (DT) learning approach to construct DT knowledge bases does not consider irrelevant and redundant attributes in problem domain. To resolve this problem, this study proposes two hybrid approaches for selecting essential attributes to construct ISS knowledge bases in SFCS. The first approach proposes an attribute selection algorithm based on the weights of artificial neural networks (ANNs) to identify the importance of system attributes. Next, using DT learning algorithm learns the whole set of training examples with important attributes to enhance knowledge learning mechanism. This hybrid ANN/DT approach is called ANN/DT-based ISS. The second approach integrates genetic algorithms (GAs) and DT learning to evolve combinatorial optimal subset of features from possible shop floor information for DT-based ISS knowledge learning mechanism. A GA is employed to search the space of all possible subsets of a large set of candidate features. For a given feature subset, a DT algorithm is invoked to generate a DT. This hybrid approach is called GA/DT-based ISS. Applying these two proposed approaches, the experiment results show that the use of essential system attributes to build scheduling knowledge bases will enhance generalization ability of the learning bias, in terms of prediction accuracy of unseen data under various performance criteria. Furthermore, simulation results indicate that the proposed approaches to build ISS improves system performance in the long run over that obtained with classical DT-based ISS and the heuristic individual dispatching rule, according to various performance criteria.