Learning for Future Mobility: Uncovering Requirements and Addressing Scalability

Technological and global trends are increasing the pace of change in mobility and infrastructure systems. At the same time, infrastructure decisions are long-lasting and often effectively permanent. Machine learning is emerging as a promising methodology to aid in the design of future infrastructure systems, due to its potential to analyze complex multi-agent dynamical systems. This talk first presents recent work on the use of deep reinforcement learning to uncover prospective engineering requirements for future mobility systems, particularly in the context of human-compatible Lagrangian traffic flow optimization. Then, several recent results are presented, which demonstrate the potential for machine learning to scale up the analysis of mobility systems, namely for multi-agent systems in grid networks and vehicle routing problems.