Developing a multi-controller by Reinforcement learning in Shop floor control system

• Main Authors: Ming-Je Cai, 蔡明哲
• Other Authors: Y. R. Shiue
• Format: Others
• Language: zh-TW
• Published: 2008
• Online Access: http://ndltd.ncl.edu.tw/handle/65256339692899725800

碩士 === 華梵大學 === 資訊管理學系碩士班 === 96 === In the current competitive environment, in order to enhance enterprise competitive advantage, manufacturing needs to control the order process of production and fast response customer demand. But in the conventional machine learning approach, such as artificial neural network （ANN）, decision tree（DT）develop an controller （i.e. scheduling controller）in shop floor control system（SFCS）to induce scheduling knowledge from a limited set of training examples. It has the main disadvantage that the classes (scheduling decision rules) to which training data are assigned must be given. However, this process becomes an intolerable time-consuming task to develop multi-controller for SFCS because of multi-decision scheduling rules for next scheduling period horizon must be determined. In addition, the local approach, whether it is based on ANN or DT learning, has the problem of global objective function （i.e. overall production performance）. That is, although the best decision rule can be selected for each scheduling controller, the combination of the selected decision rules would not simultaneously satisfy the global objective function. However, there is still very little research focusing on harmonized multi-controller for SFCS to achieve global performance criterion. To resolve above problems, this study provides a real time scheduling decision rule knowledge base（KB）to support production, multi-decision rules for next scheduling period horizon must be determined. Hence, in this study develops multi-controller to achieve these goals. The proposed multi-controller SFCS comprise five components: simulation-based training example generation mechanism, data normalization mechanism, training example clustering by two-level self-organizing map（SOM）approach, SOM-based real time decision rules class selection mechanism and the most important component: reinforcement learning（RL）-based multi-controller mechanism. In this study, the RL-based multi-controller combines SOM clustering and RL approach to select proper of decision rules in specific scheduling decision rules class based on various system statuses, current scheduling decision rules and performance criterion. Here, a number of scheduling decision rules classes are generated by two-level SOM clustering approach. Then, select proper of decision rules in specific decision rules class by proposed RL algorithm for multi-controller. Applying this approach in semiconductor FAB and flexible manufacturing system（FMS）, we expect the experiments will be demonstrate that the proposed multi-controller approach will lead to improve production performance compared to random scheduling mechanism and the heuristic individual dispatching rule based on various performance criteria.