[3K1-OS-5a] OS-5 (1)

[3K1-OS-5a-05] Signed regularization

〇Marina Takahashi¹, Hiroshi Okamoto¹, Shuji Shinohara¹, Shunji Mitsuyoshi¹, Masahiro Haitsuka², Fumihiro Miyoshi², Takatoshi Shioki², Yusuke Komine², Daisuke Ishimoto², Akihiro Tani², Yusuke Tajiri², Hidetoshi Kozono² (1. The University of Tokyo, 2. Daicel Corporation)

Keywords:Regression, Regularization, Generalization, Industrial process, Intelligent control

Our ultimate goal is to establish the intelligent control of chemical plants. Towards this, the present study puts forward regression to predict in a production process the output quality y from input factors x. Predictive control of chemical plants often necessitates learning with small datasets. Moreover, each regression coefficient must obey the sign defined by empirical knowledge of chemical plants. Here we propose to estimate the regression coefficients by regularizing their signs. The predictive performance under the signed regularization is higher than that without any regularization or under the L1 regularization. The superiority of signed regularization is more remarkable for smaller learning datasets. The signed regularization with partially wrong signs also gives higher predictive performance than the regression without any regularization. These results suggest that the signed regularization is a practical method for exploiting small learning datasets to gain the highest generalization performance.