Traffic Signal Settings Optimization Using Gradient Descent

Abstract

We investigate performance of a gradient descent optimization (GR) applied 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 depends on an activation function (e.g., TANH activation makes optimization process converge to different minima than ReLU activation).

A Comparative Analysis of Machine Learning Algorithms in Design Process of Adaptive Traffic Signal Control System

“…The authors in [15] presented a comparative analysis of optimization algorithms in optimizing phase lengths of fixed traffic signal controllers. Authors in [16] and [17] have proposed a solution to tackle the problem of congestion at an intersection by applying a decision tree classification algorithm, whereas, [18] have used gradient descent optimization technique.…”

Predicting times of waiting on red signals using BERT

“…Usually, the quality of settings can be evaluated using traffic simulations [2], but in the case of large road networks and evaluating traffic for long time periods, this method can be too time-demanding [3]. One of the recent and interesting approaches to solve this issue is to use surrogate models based on machine learning (e.g., neural networks or LightGBM) to approximate the outcomes of traffic simulations [4,5,6,7]. Such methods can be very efficient, as they return the results of evaluations a few orders of magnitude faster than in the case of computer simulations while preserving a good accuracy of approximations [4,5,6,7].…”

“…One of the recent and interesting approaches to solve this issue is to use surrogate models based on machine learning (e.g., neural networks or LightGBM) to approximate the outcomes of traffic simulations [4,5,6,7]. Such methods can be very efficient, as they return the results of evaluations a few orders of magnitude faster than in the case of computer simulations while preserving a good accuracy of approximations [4,5,6,7]. For example, the recent works showed that it may be possible to approximate the outcomes of traffic simulations where the input is a vector representing signal settings and the output computed by the simulations is the total time of waiting on red signals in a given urban area and in a given time period [4,5,6,7].…”

“…Such methods can be very efficient, as they return the results of evaluations a few orders of magnitude faster than in the case of computer simulations while preserving a good accuracy of approximations [4,5,6,7]. For example, the recent works showed that it may be possible to approximate the outcomes of traffic simulations where the input is a vector representing signal settings and the output computed by the simulations is the total time of waiting on red signals in a given urban area and in a given time period [4,5,6,7]. However, these works assumed that durations of green and red signal phases are constant (e.g., lasting 58 and 62 seconds, respectively [4]) and the only possible way of controlling traffic signal settings is by modifying traffic signal offsets (time in seconds from the beginning of a simulation to the first switch from the red signal state to the green signal state for a representative signal from a given intersection).…”

“…In this paper, we evaluate 2 BERT-based models and compare them with other machine learning models which gave good results in the previous research works [4,5,6,7] (feed forward neural networks, graph neural networks, graph convolutional neural networks, LightGBM), which are summarized in Section 2. We ran a series of experiments on a large dataset generated using traffic simulations on a realistic road network.…”