Accelerated Training for Connected and Automated Vehicles Based on Adaptive Evaluation Method

Principal Investigator(s):

Henry Liu, Director – Center for Connected and Automated Transportation (CCAT)
Director – Mcity
Professor of Civil and Environmental Engineering – The University of Michigan
Research Professor – The University of Michigan Transportation Research Institute

Project Abstract:
This project focuses on resolving the inefficiency problem caused by the long-tail phenomena in the development of connected and automated vehicles (CAVs) to accelerate the training process. The training of CAV model can be divided into two stages: in the first stage, the model is trained with naturalistic driving data, while in the second stage, when the training efficiency is greatly compromised by the long-tail phenomena, a reinforcement learning-based mechanism with critical scenarios is proposed. The critical scenarios, which contain vulnerabilities of the CAV model, can be generated by the adaptive evaluation method. An incremental learning mechanism is designed and a discount factor will be introduced according to the probability of the critical scenarios. Importance sampling technologies will be applied to guarantee the accuracy of the discount factor. Meanwhile, a training and testing platform will be designed and built to validate the proposed accelerated training framework.

Institution(s): University of Michigan Transportation Research Institute

Award Year: 2019

Research Thrust(s): Control & Operations, Enabling Technology, Modeling & Implementation

Project Form(s):