請用此 Handle URI 來引用此文件:
http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/91507
標題: | 政策學習於製程參數:以因果推論觀點 Policy Learning in Manufacturing Process Parameters: A Causal Inference Perspective |
作者: | 陳柏儒 Bo-Ru Chen |
指導教授: | 李家岩 Chia-Yen Lee |
關鍵字: | 製程參數最佳化,個體行為指派,政策學習,政策詮釋, Process Parameter Optimization,Individual Action Assignment,Policy Learning,Policy Interpretation, |
出版年 : | 2023 |
學位: | 碩士 |
摘要: | 本研究探討製程參數最佳化之問題,在一製造系統中可控因子往往會直接影響到產品品質。然而,除了可控因子之外仍有其他不可控因子會影響到生產環境,進而影響到產出,而這些不可控因子通常與每個決策個體有關,同時代表著該個體的特性。在過去的製程參數最佳化研究當中,實驗設計常用以獲得最好的製程參數設定,同時也有大量的研究利用預測模型搭配最佳化方法求解,然而個體的異質性鮮少被當作決策依據。因此,我們透過觀察性資料進行因果推論中的雙重穩健估計同時衡量可控因子與不可控因子對於產出的影響。在我們所提出的架構當中涵蓋了對於資料的前處理、政策品質估計、政策最佳化與政策詮釋,給定一個體,所習得的政策便會基於其個體特性給出製程參數設定值,又由於此政策是一決策樹結構,因此其自有高度的可詮釋性。在數值研究當中,我們所提出的方法展現出良好的決策能力,同時透過實驗我們也建議使用者採用政策森林,以加強該方法之決策品質,而政策森林不足之可解釋性也可透過我們所提出之政策詮釋方法加以補強。據我們所知,此研究率先提出一特定方法來詮釋習得政策,並且也以一容易理解的方式來呈現決策準則。一決策者可透過習得之政策基於每個個體特性給予其最佳的製程參數設定,同時透過決策之詮釋也可獲得製程背後之洞見。 We study the process parameter adjusting problem to maximize a quality quantity, while considering the heterogeneity of instances. For a manufacturing process, controllable factors such as process parameters will directly affect the quality measurement. However, there are also some uncontrollable factors, which are usually instance-wise features, influencing the outcome. In previous studies, conducting an experiment or developing prediction models to select optimal controllable factors are well-known methods. However, they lack consideration of the effect of uncontrollable factors, i.e., heterogene- ity of instances. In this study, we use observational data and adopt doubly robust esti- mation in causal inference to estimate the effect of both controllable and uncontrollable factors. Our proposed framework involves data preprocessing, policy value estimation, policy optimization, and policy interpretation. The learned policy inputs an instance-wise feature and outputs an action to adjust process parameters. This policy is also highly interpretable due to its tree structure. In the numerical study, our method shows great support for decision-making. We also recommend taking Policy Forest as the policy class to increase the quality of a solution. Meanwhile, the insufficient interpretability can be enhanced by our policy interpretation procedure. To the best of our knowledge, this is the first study to interpret a learned policy and present decision rules in a more understand- able fashion. With the learned policy, a decision maker can determine the best process parameter setting for each individual according to its characteristics. Moreover, insights about the process can also be obtained via policy interpretation to drive productivity. |
URI: | http://tdr.lib.ntu.edu.tw/jspui/handle/123456789/91507 |
DOI: | 10.6342/NTU202303926 |
全文授權: | 同意授權(全球公開) |
顯示於系所單位: | 資訊管理學系 |
文件中的檔案:
檔案 | 大小 | 格式 | |
---|---|---|---|
ntu-111-2.pdf 此日期後於網路公開 2026-08-31 | 2.2 MB | Adobe PDF |
系統中的文件,除了特別指名其著作權條款之外,均受到著作權保護,並且保留所有的權利。