2297. Multi-criteria Traffic Signal Control Optimization using Multi-policy Reinforcement Learning
Invited abstract in session TA-59: Road network optimization, monitoring and control, stream Transportation.
Tuesday, 8:30-10:00Room: Liberty 1.14
Authors (first author is the speaker)
| 1. | Zakir Farahmand
|
| Civil Engineering and Management, University of Twente | |
| 2. | Oskar Eikenbroek
|
| Transport Engineering and Management, University of Twente | |
| 3. | Konstantinos Gkiotsalitis
|
| Civil Engineering, NTUA | |
| 4. | Eric van Berkum
|
| Civil Engineering, University of Twente |
Abstract
With increasing mobility demand, a key challenge in urban areas is to efficiently allocate limited road capacity among different user-groups, including passenger cars, freight, public transport, bicycles and pedestrians. Particularly at intersections with demand-supply imbalance, poor traffic signal controls (TSC) leads to excessive delays, increased fuel consumption, higher emissions, noise pollution and air quality deterioration. Conventional TSCs like vehicle-actuated typically focus on minimizing waiting time, often conflicting with reducing emissions arising from frequent braking and acceleration manoeuvres, especially for heavy-duty vehicles. The primary question is how to balance between conflicting objectives while responding to highly volatile traffic demand in real-time? In this study, we propose a novel modular multi-policy deep reinforcement learning (MMDRL) in which the agent aims to find a set of optimal policies. The method combines the theory of Pareto optimality and Deep Q-Networks to derive a Pareto front of non-dominated policies without relying on a priori preferences or scalarization function. This enables the agent to simultaneously minimize (i) person-based delays and (ii) emissions (CO₂, CO, NOₓ) near intersections. Simulation results show that MMDRL significantly outperforms conventional (e.g., vehicle-actuated) TSCs in waiting times, emissions, and number of stops while dynamically adjusting signal plans to changing traffic conditions.
Keywords
- Programming, Multi-Objective
- Programming, Dynamic
- Transportation
Status: accepted
Back to the list of papers