Research

Multi-agent Graph Reinforcement Learning for Connected Automated Driving

[video] [paper] [code]

McGill University, Jan. 2020~ present
Advisors: Prof. Lijun Sun

  • Utilize the graph attention networks in the navigation setting of multi-agent reinforcement learning for mixed-autonomy cooperation.
  • Introduce Dynamic relational index based on both velocity and position information to capture attention features among surrounding agents.
  • Conduct extensive experiments in different difficult levels of transportation networks against different baselines to demonstrate the effectiveness of the proposed approach..

Efficient Motion Planning for Automated Lane Change based on Imitation Learning and Mixed-Integer Optimization

[video] [paper] [code]

  • Transformed the lane change mission into Mixed Quadratic Problem (MIQP) with logical constraints to guarantee safe and comfortable lane change movements.
  • Proposed a hierarchical imitation learning with “classification layer” and “action generation layer” to provide online, fast and more generalized motion planning.

Optimizing Control Performance based on Deep Neural Network

[video] [paper] [code]

University of California Berkeley, June. 2018~ September. 2018
Advisors: Dr. Ching-Yao Chan and Dr. Pin Wang

  • Proposed a new methodology which combine Principal Component Analysis (PCA) and Time Delay Neural Network to evaluate automated vehicle control performance and model the behavior of low level controller.
  • Designed an optimized feed-forward compensator based on deep neural network and achieved improved performance in U-turn scenario.

Research on Decision-making and Control System based Deep Reinforcement Learning

[video] [paper] [code]

University of California Berkeley, June. 2018~ September. 2018
Advisors: Dr. Ching-Yao Chan and Dr. Pin Wang

  • Designed two similar Deep Q learning frameworks with quadratic approximator for deciding how to select a comfortable gap and follow the preceding vehicle.
  • Proposed a novel hierarchical deep reinforcement learning for decision making and control of lane change situations.