Building Intersection Digital Twins via GPU-Accelerated Human Regularized Reinforcement Learning

This work tackles a key limiting factor in the design of congestion reduction schemes, namely, the challenge of validating a proposed scheme with sufficiently realistic actor behavior. This is critical as oversimplified driving behavior models could artificially inflate or reduce the impact of a proposal. This project aims to tackle this challenge using new techniques in artificial intelligence and simulator design to enable the design of realistic driver models. A two-stage approach is proposed to design an efficient simulator to acquire the large number of samples needed to learn more complex models of human behavior at intersections. The team will use Goal-Conditioned Multi-Agent Reinforcement Learning (GC-MARL), combined with imitation learning, to build competent but diverse models of human behavior in dense traffic settings. By combining the agents and the simulation capabilities, the team will construct an open-source platform that potential state and federal partners can use to evaluate congestion mitigation schemes in realistic settings. Users will be able to provide different road layouts, speed limits, traffic control settings, etc., and be able to quickly evaluate the performance with diverse agents capable of non-idealized behaviors such as red light running, inching into intersections, performing cut-offs of agents in nearby lanes and other complex, realistic behaviors. As an additional benefit, this proposal will sharply reduce the cost needed to apply reinforcement learning for congestion reduction.


  • English


Subject/Index Terms

Filing Info

  • Accession Number: 01897924
  • Record Type: Research project
  • Source Agency: Connected Communities for Smart Mobility Towards Accessible and Resilient Transportation for Equitably Reducing Congestion (C2SMARTER)
  • Contract Numbers: 69A3551747124
  • Files: UTC, RIP
  • Created Date: Oct 30 2023 10:29PM