Gradient-based optimization has been extremely successful, especially in machine learning, efficiently scaling to models with billions of parameters. I discuss my research on creating a differentiable traffic simulator with the goal of making traffic optimization easier. There are several challenges in differentiating a practical, discretized, traffic microsimulation, owing to the inherent discontinuities and lack of sensitivity (e.g. lane changing, state/parameter-dependent switching conditions). I explain sufficient conditions for the objective to be (lipschitz) continuous and illustrate the implications and future research directions with several interactive examples.
Back to Mathematical Challenges and Opportunities for Autonomous Vehicles