Deep reinforcement learning-based multi-lane mixed traffic ramp merging strategy.

Area of Science:

• Intelligent Transportation Systems
• Traffic Engineering
• Reinforcement Learning in Traffic Control

Background:

• On-ramp merging is critical for hybrid traffic control but current methods often overlook multi-lane coordination.
• Existing research primarily optimizes single-lane ramp merging, neglecting complex multi-lane traffic dynamics.

Purpose of the Study:

• To develop an advanced on-ramp merging strategy for multi-lane traffic flow.
• To enhance vehicle coordination and safety during merging maneuvers.

Main Methods:

• Proposed an Improved Dueling Double DQN (D3QN) On-ramp Merging Strategy (IDS) integrated with a sine function.
• Established a Vehicle Coordination System (VCS) to guide multi-lane merging.
• Implemented an action masking mechanism to ensure safe exploration during strategy development.
• Validated the strategy using SUMO simulations under varying traffic densities.

Main Results:

• The VCS + IDS strategy achieved a 98.62% merging completion rate and 98.11% task completion rate at 1200 vehicles/lane/hour.
• Demonstrated significant improvements of 11.08% and 10.79% in completion rates compared to traditional D3QN.
• Effectively guided vehicles to find suitable gaps in traffic flow, enhancing safety and efficiency.

Conclusions:

• The proposed VCS + IDS strategy is effective for optimizing on-ramp merging in multi-lane traffic.
• The integration of D3QN and sine function provides a robust approach to traffic coordination.
• This method offers a significant advancement over traditional D3QN in handling complex merging scenarios.