student and faculty conducting research

PhD Students and Post-Doctoral Researchers

Juan C. Medina

  • Advisor:
      • Rahim F. Benekohal
  • Departments:
    • Civil and Environmental Engineering
  • Areas of Expertise:
      • Agent-based Control
      • Connected Vehicles
      • Transportation Safety
      • Intelligent Transportation Systems
      • Traffic Control Systems

  • Thesis Title:
      • MASTraf: a decentralized multi-agent system for network-wide traffic signal control with dynamic coordination
  • Thesis abstract:
      • From traditional pre-timed isolated intersections to actuated and coordinated corridors, traffic signal control in urban networks has evolved into complex adaptive systems that mitigate effects of increasing traffic demands with limited available capacity. However, unexpected or rapid changes in volumes, oversaturation, incidents, and adverse weather conditions, among others, significantly impact traffic operations in ways that adaptive signal systems cannot always cope with. MASTraf presents a new paradigm in traffic signal control that not only adapts to traffic conditions in real time, but that is also performance-driven, capable of self-adjusting control policies to optimize a set of objectives. For a signalized intersection, a controller (or agent) is proposed to sense the surrounding state of traffic, take actions, and assess the effects of such actions to create control policies using machine learning (e.g. Q-learning and approximate dynamic programming). MASTraf is proposed to reduce the complexity of the problem, where independent and cooperative learning agents at each intersection engage in asynchronous message passing of local information to assess the potential outcomes of their actions, biasing action selection to improve both local and system-wide performance. The combination of machine learning and the cooperative algorithm yields emergent agent (or traffic signal control) coordination that resembles the behavior of current signal control strategies and is also capable of self-improving performance. MASTraf was implemented in simulated scenarios ranging from a single intersection, to arterials and grid networks. Results show that decentralized and cooperative agents not only provided signal coordination and efficient capacity utilization, but also prevented queue spillbacks and gridlocks even in highly congested conditions. When the proposed MAS is coupled with an approximation of the state of traffic using a functional form, the system becomes well-suited for high dimensional state representations that can incorporate enhanced sensor data from connected vehicles, weather, incidents, and virtually any aspect that may impact the traffic stream.
  • Downloads:

Contact information: