This project develops a transformer-based framework for personalized promotion delivery under dynamic budget constraints. By leveraging autoregressive sequence modeling and generative pretraining techniques, it overcomes the scalability and retraining limitations of conventional CMDPs and Q-learning. Key innovations include dynamic adjustment of user cost thresholds and self-attention-based credit assignment to identify high-impact user segments.

This work introduces a state-dependent framework for dynamic contest design. Integrating behavioral models of user selection, demand forecasting, and real-time event triggers, the system adaptively modifies contest offerings to optimize engagement and mitigate cannibalization. The project combines robust stochastic programming with reinforcement learning for live optimization under high-dimensional constraints.

This project extends Pulse Wave Imaging (PWI), a high-frame-rate ultrasound technique, to both central and peripheral arteries, enabling real-time, non-invasive vascular health monitoring during exercise. It integrates unsupervised deep learning for elasticity and pressure estimation and is advancing toward miniaturized, wearable ultrasound sensors using PMUT arrays.

This research leverages a robotics testbed (RobUST) to improve motor skill acquisition through assist-as-needed training. Using real-time EMG, EEG, motion capture, and force plate data, the system characterizes expert-novice motion differentials and applies machine learning to generate adaptive, personalized force feedback profiles for movement correction and retention.

This project formulates a data-driven, behaviorally aware optimization pipeline to manage contest recommendation and user retention strategies. It explicitly incorporates peak experience effects—sustained engagement after single wins—into RL-based engagement maximization, balancing short-term actions with long-term user value.