Differentiable Modeling of Traffic Signals for Gradient-based Parameter Optimization in Microscopic Traffic Simulations

Beier, Mathes (2022) Differentiable Modeling of Traffic Signals for Gradient-based Parameter Optimization in Microscopic Traffic Simulations. Bachelor thesis, Institute for Visual and Analytic Computing, University of Rostock.

Text Bachelorarbeit_Mathes_Beier.pdf Restricted to Registered users only Download (682kB)

Abstract

In many cases traffic signals are getting optimized by simulation based methods. Since detailed simulations are mostly optimized without partial derivatives, they tend to act in the way of a black-box. Thus the optimization method represents a black-box-procedure. In this thesis we want to transform a microscopic traffic intersection in a differentiable model to employ a gradient-based optimization method on our traffic model to optimize it. The difficulty lies in the ability to differentiate our traffic simulation in every step of our traffic simulation. Thus we have to replace hard branching methods by a smoother version. In this thesis we will formulate the problem by making use of a timed automaton and limit our self on a single traffic intersection with valid traffic signal timings. The goal is to develop a microscopic traffic simulation with the help of a classic car-following model and transfer it to a smoother version which is differentiable in every operation. Following that we will continue to utilize automatic differentiation methods to calculate gradients which we can apply to our optimization to find the local optimum. By testing our optimization on 100 scenarios with random distributed traffic loads we found out, that in a simple evaluation scenario, our method converges very closely to the global optimum.