Traffic Signal Settings Optimization Using Gradient Descent

We investigate performance of a gradient descent optimization ap- plied to the traffic signal setting problem and compare it to genetic algorithms. We used neural networks as metamodels evaluating quality of signal settings and discovered that both optimization methods produce similar results, e.g., in both cases the accuracy of neural networks close to local optima depend on an activation function (e.g., TANH activation makes optimization process to converge to different minimas than ReLU activation).