个人信息
参与实验室科研项目
基于强化学习的呼吸机参数智能精准控制技术研究
学术成果
共撰写/参与撰写专利 0 项，录用/发表论文 2 篇，投出待录用论文1篇。
Conference Articles
-
RLMV: An Algorithm for Predicting the Duration of Mechanical Ventilation in Patients
Jiaying Xi,
Yun-Bo Zhao ,
Haoquan Zhou,
Siyuan Guo,
and Chenwei Xu
In
2026
[Abs]
[pdf]
Mechanical ventilation is a supportive therapy for patients with respiratory failure, both inadequate and prolonged ventilation pose significant health risks, including increased mortality. However, the optimal duration of mechanical ventilation is difficult to predict and relies heavily on real-time clinical assessment. To address the challenge, we propose a novel predictive algorithm termed RLMV. Specifically, we represent the first attempt to integrate static physiological features with dynamic temporal variations, constructing a cascaded prediction model that combines “Feature Regression Screening” with “Temporal Deep Modeling”. Experimental results based on the MIMIC-III dataset demonstrate the superiority of RLMV, which achieving a Mean Squared Error (MSE) of 0.0032 and a Mean Absolute Error (MAE) of 0.0324 and outperforming the existing methods.
Journal Articles
-
Human-in-the-Loop Reinforcement Learning with Risk-Aware Intervention and Imitation
Yaqing Zhou,
Yun-Bo Zhao ,
Chenwei Xu,
Chen Ouyang,
and Pengfei Li
Expert Syst. Appl.
2026
[Abs]
[doi]
[pdf]
Human-in-the-loop reinforcement learning (HiL-RL) improves policy safety and learning efficiency by incorporating real-time human interventions and demonstration data. However, existing HiL-RL methods often suffer from inaccurate intervention timing and inefficient use of demonstration data. To address these issues, we propose a novel framework called HiRIL (Human-in-the-loop Risk-aware Imitation-enhanced Learning), which establishes a closed-loop learning mechanism that integrates risk-aware intervention triggering and imitation-based policy optimization under a dual-mode uncertainty metric. At the core of HiRIL is the Bayesian Implicit Quantile Network (BIQN), which captures both epistemic and aleatoric uncertainty through Bayesian weight sampling and quantile-based return modeling. These uncertainties are combined to generate risk scores for state-action pairs, guiding when to trigger human intervention. To better utilize intervention data, HiRIL introduces a prioritized experience replay mechanism based on risk difference, which emphasizes human interventions that significantly reduce risk. During policy optimization, a local imitation loss is applied to clone human actions at intervention points, enabling risk-guided joint optimization. We conduct extensive experiments on the CARLA end-to-end autonomous driving benchmark. Results show that HiRIL consistently outperforms baselines across multiple metrics and maintains strong robustness under perturbations and non-stationary human intervention.