欧阳晨同学论文被《Expert Systems With Applications》接受发表
欧阳晨 周雅情 赵昀昇 题为 “Using Non-Intervention as Information: Constraint-Based Online Reward Learning from Human Feedback” 的论文已被《Expert Systems With Applications》接受发表。该论文摘要如下:
Intelligent machines deployed in real-world environments inevitably make mistakes where humans intervene and provide valuable corrective feedback. Many existing works focus on learning reward from such corrective trajectories, while largely overlooking the non-intervention stage. However, the non-intervention stage is argued to convey meaningful and complementary information about the underlying reward. Based on this insight, we propose a constraint-based framework for online reward learning from human feedback, which explicitly exploits information from both intervention and non-intervention stages. Under a bounded suboptimality assumption, human feedback from these stages is interpreted as inducing stage-dependent relaxed optimality constraints on the reward, which together define a feasible set of reward hypotheses consistent with observed behavior. To make this formulation tractable in practice, the feasible set is approximated by a polyhedral outer approximation constructed from sampled trajectories, and a representative reward parameter is obtained through log-barrier estimation over the resulting constraints. Experiments in representative simulation tasks and a preliminary real-world robotic manipulation study show that our approach yields more accurate and stable reward learning performance compared to existing intervention learning methods.