Search Papers | Poster Sessions | All Posters
Poster A36 in Poster Session A - Tuesday, August 6, 2024, 4:15 – 6:15 pm, Johnson Ice Rink
Modeling Surprise in a Physically Grounded Joint Inference of Preference, Knowledge, and Perceptual Access
Harry Chen1 (), Karen Chung1, Abhishek Bhandwaldar2, Joshua B. Tenenbaum1, Tomer D. Ullman3, Tianmin Shu4; 1Massachusetts Institute of Technology, 2MIT-IBM Watson AI Lab, 3Harvard University, 4Johns Hopkins University
Even infants understand other agents can have partial observability of the world, and show varying degrees of uncertainty about the knowledge and preferences of others. This work models people's inference of another agent's preference and knowledge given limited perceptual access, as measured by their surprise response. We propose POPIK (Physically-grounded Observation, Preference, and Knowledge Inference), a Bayesian inverse-planning method that models graded surprise in the inference of preference, knowledge, and perceptual access in rich 3D environments. To test our model, we extended the AGENT dataset to trials that probe preference, knowledge, and perceptual access. Experimental results show POPIK replicates humans' varying degrees of surprise when judging the behavior of agents under different visibility conditions. These results suggest that reasoning about how agents plan in imagined physical states according to their knowledge under limited observability is key to reverse-engineering human-like uncertainty judgments in psychological reasoning tasks.
Keywords: intuitive psychology perceptual access knowledge Bayesian inverse planning