Training neural networks to approximate traffic simulation outcomes

Abstract

We present results of our research on training neural networks to approximate traffic simulation outcomes, such as total times of waiting on a red signal. We developed TensorTraffic software, based on a TensorFlow library, and trained neural networks on a dataset generated by simulating traffic on a realistic road network of Warsaw using Traffic Simulation Framework software. The goal of conducted experiments was to approximate the total times of waiting on a red signal on a region of Warsaw (Stara Ochota district), with the input to neural nets representing offsets of traffic signals on that region. In the presented research, we focused on investigating different neural network models and strategies of their training. We took into account different sizes of training sets, different numbers of neurons and layers, different parameters of dropout and learning rate, in order to reduce as much as possible time required to conduct experiments and apply this method in practice (e.g., time required to generate training and test sets for neural networks, time to train neural networks and time to make inferences on a new set), while preserving a sufficient accuracy of approximations, which may be especially important from a practical point of view Results show that it is possible to train neural networks able to approximate with a high accuracy (with an average error 1.18%) outcomes of traffic simulations. Moreover, TensorTraffic allows obtaining results of approximations a few orders of magnitude faster than by running simulations using microscopic traffic models and it is possible to achieve acceptable accuracy on a relatively small training set (consisting of 10240 elements). It means that the method can be potentially applied to different traffic analysis and transport planning tasks (e.g., to find suboptimal configurations of traffic signals) and time-consuming computer simulations applied nowadays for traffic analysis can be potentially replaced by neural nets supported by computations using graphical processing units (GPU). This innovation may significantly reduce time required to complete research and engineering tasks related to designing road infrastructure and analysing vehicular traffic, as well as enable developing better traffic management systems. As an example, we applied the method to the traffic signal setting problem and accelerated existing genetic algorithm, giving opportunity to evaluate much larger set of possible settings and find better traffic management strategies.