A Dynamic Generator for Machine Learning Training for Traffic Management Systems

Abstract

This study introduces the concept of a Dynamic Generator for Machine Learning (ML) training. The paper presents the use of this key concept in order to obtain valuable data about the traffic in a crowded city like Bucharest according to the city habits and to the real-time data collected. The trained ML models may be used in the architecture of a modern Traffic Management System (TMS) based on Artificial Intelligence (AI) for Bucharest or any other crowded city. The Generator consists of a Simulator and a Collector of real data from the observed environment. In this way, the obtained set of data is hybrid because it may contain real data obtained from sensors together with synthetic data obtained from the Simulator. The Generator is dynamic because the synthetic data are smoothly replaced with real data as time goes by and as more and more sensors are put into operation in the city. The Simulator results are produced according to a simplified city model and some socio-compartmental parameters which try to describe at the macro level the behavior of the city based on historical data collected in time by the authorities and other nongovernmental organizations. These initial parameters could be improved, and more than that, the city's behavior is changing over time. That's why the city model and the parameters have to be refined in time. In this way, each new data generation will serve to re-train more accurately the set of ML models needed in the Traffic Management System.

Keywords

• informatics management system
• simulation
• artificial intelligence
• machine learning
• traffic