Building Human-Aware AI: Learning from and Assisting Human Decision-Makers

Engineering notes

Key topics: autonomous driving, prediction. See the paper for implementation details and experimental results.

Chinese explanation / 中文解读

中文解读待补充：本站会优先为端到端自动驾驶、BEV感知、3D目标检测、轨迹预测、路径规划、LiDAR感知等高价值论文补充中文说明。

Original abstract

As artificial intelligence systems become more capable, they are increasingly used not only as standalone problem-solvers, but as systems that learn from and interact with humans. In these settings, success depends not only on the strength of the model in isolation, but also on how well it fits the humans it is trained on or deployed alongside. This dissertation argues that human heterogeneity is a central ingredient in the design of effective human-aware machine learning systems. Rather than treating differences between people as noise to be averaged away, I show that variation in human expertise and preferences can provide useful structure for both learning and assistance. I study this question through three lines of work. First, I examine behavioral cloning from human demonstrations and show that, in small-data regimes, low-skill data can sometimes produce stronger learned policies than higher-skill data. I explain this through differences in the fragility of demonstrator behavior, show how demonstrator skill can be leveraged in curriculum design, and show that the resulting models remain ``human-like". Second, I study AI assistance in sequential decision-making, using chess as a case study. I show that effective interventions should account not only for the best action in isolation, but for how the assisted human is likely to act afterward. By modeling both human policy and human value, I develop intervention strategies that better support downstream human performance, and validate my approach both in simulation and in a user study with real human chess players. Third, I study human-AI interaction in ethical decision-making through a kidney allocation experiment. I show that humans are more likely to rely on AI recommendations when the AI's exhibited values demonstrate alignment with the human's values, but not when the AI simply claims value similarity. I also find that AI predictions can have downstream effects on human ethical preferences, and that these effects are stronger than human predictive assistance. Together, these results show that modeling human heterogeneity can improve how AI systems learn from people, assist them, and influence their decisions.

Links and sources

• PDF from original source

Comments