可见的
强化学习
缺少数据
控制(管理)
钢筋
计算机科学
人工智能
控制理论(社会学)
机器学习
心理学
物理
社会心理学
量子力学
作者
Haiyan Yu,Jiaojun Zhu,Nan Kong,Li Luo,Ching-Chi Yang
摘要
In this era of artificial intelligence, industrial engineering can benefit greatly from a reliable model for system dynamics when dealing with sequential decision problems under uncertainty, yet a biased model can yield deceptive and sometimes disastrous outcomes. Acknowledging data quality is critical to training a reliable model, incomplete data in many situations often poses a significant challenge to maintaining data quality, which in turn can affect the decision quality. In this paper, we investigate the use of a partially observable reinforcement learning (PORL) approach to handle unevenly spaced missing data for sequential decision-making under uncertainty. With a given batch of tabular data, this PORL-based algorithmic approach uses the Metropolis- Hastings (M-H) algorithm for incomplete-data imputation and low-discrepancy sequences to reduce bias in the imputation. The PORL method then employs the training data after missing imputation to learn the optimal policy. We assess its efficacy numerically with different percentages of missing data and show our imputation technique outperforms the mean imputation and standard M-H algorithms in the PORL framework. We further showcase the usability of our smart algorithm in the application of personalized blood glucose control, which is expected to improve the robustness of automated insulin delivery for diabetes care. The proposed analytics method can lead to the development of an efficient tool for making medication decisions (even reducing the mean absolute relative difference) under missing data from CGM.
科研通智能强力驱动
Strongly Powered by AbleSci AI