Projects
⭐ Imitating Cost Constrained Behaviors in Reinforcement Learning
We explore imitation learning in scenarios where expert behavior is influenced by both rewards and constraints, introducing methods such as Lagrangian, meta-gradient, and cost violation-based approaches to address trajectory cost constraints, with empirical results showing that the meta-gradient-based approach outperforms existing methods in accurately imitating cost-constrained behaviors
⭐ Preference-Aware Delivery Planning for Last-Mile Logistics
Our research addresses the challenge of optimizing last-mile logistics delivery routes, proposing a hierarchical route optimizer with learnable parameters that integrates optimization and machine learning to bridge the gap between optimized routes and practitioner-preferred routes, which often arise from differing priorities.
⭐ LLM-based Early Rumor Detection with Imitation Agent
Our research tackles Early Rumor Detection (EARD) by predicting the earliest point to assess a claim's truthfulness from social media. A lightweight agent analyzes time-series data while an LLM detects rumors, using a Markov Decision Process (MDP) with expert trajectories.