Development of Learning Agent-based Systems for Improved Urban Traffic Control (DLASIUT)
Funding body: Croatian Science Foundation (HRZZ)
Budget: 1,088,840.00 HRK (143,268.00 EUR)
Project duration: 15th February 2021 – 14th February 2025
Project leader: Edouard Ivanjko, Ph.D.
Today’s urban environments are prone to daily congestions due to dense traffic. The development of Machine Learning (ML) based traffic control systems for such environments gathered interest to create intelligent systems with the potential to improve the existing transport network efficiency. Applying ML in control of complex urban environments is prone to the curse of dimensionality. The controller behavior is influenced by the number of observed traffic parameters describing the environment in which it acts. Rising the number of parameters causes an exponential increase of possible state-action space, making it nearly impossible to find an optimal control policy in a reasonable time. The scalability of the same space becomes very important. It is also necessary to gain trust or confidence in the ML control system’s performance in unforeseen situations. Having a control policy that performs well in all relevant traffic states is more important than superior performance in some states. Thus, tuning of such systems for significant transport demand changes is very problematic or even infeasible for operators without computer support. The main benefit of the project DLASIUT is the proposed learning framework and structure of an agent-based traffic controller capable of learning the optimal control policy from microscopic simulations containing realistic models of a real-world urban road network. Additionally, support to Connected and Autonomous Vehicles (CAVs) will be added using them as an extra control output ensuring the applicability in future mixed traffic flows containing classic vehicles and CAVs. In-depth testing using realistic simulation models and Structured Simulation Framework from transport technology point of view to identify possible poor controller behavior will improve the state of the art of ML-based traffic controllers. The benefit for the citizens of urban environments is better traffic management and reduction of congestions and vehicle emissions.