Multi-Lane Differential Variable Speed Limit Control via Deep Neural Networks Optimized by an Adaptive Evolutionary...

Area of Science:

• Intelligent Transportation Systems
• Machine Learning
• Control Theory

Background:

• Variable speed limits are essential for advanced traffic management.
• Deep reinforcement learning (DRL) faces challenges with delayed rewards and convergence in traffic control.
• Evolutionary strategies offer robust optimization for complex control problems.

Purpose of the Study:

• To develop a novel approach for multi-lane differential variable speed limit control.
• To address the limitations of traditional DRL in traffic management, specifically delayed rewards and convergence instability.
• To enhance freeway throughput and reduce emissions using intelligent control strategies.

Main Methods:

• Utilizing Covariance Matrix Adaptation Evolution Strategy (CMA-ES), a gradient-free optimization method.
• Employing a deep-learning-based model for dynamic, lane-specific speed limit learning.
• Optimizing neural network parameters via a multivariate normal distribution with a dynamically updated covariance matrix based on freeway throughput.

Main Results:

• The proposed CMA-ES method significantly outperformed DRL, traditional evolutionary methods, and no-control scenarios in simulations.
• Achieved a 23% improvement in average travel time.
• Demonstrated an average 4% reduction in CO, HC, and NOx emissions, with explainable speed limits and strong generalization.

Conclusions:

• The CMA-ES based approach effectively manages variable speed limits in complex traffic scenarios.
• This method offers a robust and efficient alternative to traditional DRL for traffic control, improving efficiency and environmental impact.
• The approach provides explainable control actions and generalizes well to different traffic conditions.