Simulation Study on a Three-Segment Highway Network Found that Artificial Intelligence (AI) Based Ramp Metering Reduced Average Time Spent by Vehicles by 35 Percent.
Simulations Performed on a Benchmark Small-Scale Highway Network Reduced Network Congestion.
Netherlands
Reinforcement Learning With Model Predictive Control for Highway Ramp Metering
Summary Information
Ramp metering is widely used as an effective traffic operations management tool to reduce overall congestion by regulating the traffic flow at the freeway onramps. Ramp metering can be done via fixed timings based on historic traffic data or dynamic timings that respond to real-time traffic conditions. This study aimed to use a method that combined a traditional ramp metering control algorithm with an AI-based model for ramp metering to optimally balance mainline freeway congestion and queue lengths at the onramp.
METHODOLOGY
The study combined Model Predictive Control (MPC), a traditional control approach that optimizes decisions over a short future time horizon, and an AI algorithm, Reinforcement Learning (RL), which automatically learned from data a policy that optimally balanced mainline traffic congestion and queue lengths at on-ramps, without extensive open-loop data for further manual tuning. A macroscopic simulation program was used for the analyses, in which a hypothetical three-kilometer freeway with one onramp was generated. A constrained ramp metering control scenario was defined by imposing a maximum queue length of 50 vehicles at the onramp. The proposed MPC-RL controller was tested in comparison to three other ramp metering control policies, and the results were averaged over 15 simulation runs.
FINDINGS
- The MPC-RL controller decreased the time spent by vehicles in the network on average, from around 1,100 vehicle hours to 700 vehicle hours (a 35 percent reduction).
- In contrast, the other three state-of-the-art control approaches had, respectively, 1) more total time spent and more constraint violations, 2) more total time spent and more control variability, and 3) variability and violations decreasing at a lower rate. The proposed approach was robust to the variability of the randomly generated demand and congestion scenarios and could avoid all constraint violations.
