A0832
Title: Diffusion policies with offline inverse reinforcement learning to promote physical activity in older adults
Authors: Chang Liu - University of Central Florida (United States)
Ladda Thiamwong - University of Central Florida (United States)
Yanjie Fu - Arizona State University (United States)
Rui Xie - Universify of Central Florida (United States) [presenting]
Abstract: Offline reinforcement learning (RL) holds promise for healthcare applications but faces key challenges, including defining rewards and managing distributional shifts between learned policies and human behavior. To address these issues in promoting physical activity among older adults at high fall risk using wearable sensor data, KANDIKolmogorov-Arnold Networks and Diffusion Policies for Offline Inverse RL are introduced. KANDI leverages the powerful function approximation of Kolmogorov-Arnold Networks to learn reward functions from low-risk (expert) behavior and uses diffusion-based policies within an Actor-Critic framework to refine actions and mitigate distributional discrepancies. Evaluated on data from a two-arm clinical trial, KANDI identifies optimal intervention timing tailored to fall-risk levels, effectively promoting daily activity. Synthetic experiments show KANDI outperforms existing RL methods in both reward inference and policy performance, demonstrating its practical value in real-world health interventions.
