[Paper Review] Cooperative Driving at Unsignalized Intersections Using Tree Search

Huile Xu, Yi Zhang+2 more|arXiv (Cornell University)|Feb 4, 2019

Traffic control and management19 references17 citations

TL;DR

This paper proposes a cooperative driving strategy for connected and automated vehicles (CAVs) at unsignalized intersections using Monte Carlo Tree Search (MCTS) enhanced with heuristic rules to find a nearly global-optimal passing order within milliseconds. The method achieves near-optimal delay reduction—demonstrated by 80% lower average delay than FIFO—while maintaining real-time computation performance through intelligent tree exploration and pruning.

ABSTRACT

In this paper, we propose a new cooperative driving strategy for connected and automated vehicles (CAVs) at unsignalized intersections. Based on the tree representation of the solution space for the passing order, we combine Monte Carlo tree search (MCTS) and some heuristic rules to find a nearly global-optimal passing order (leaf node) within a very short planning time. Testing results show that this new strategy can keep a good tradeoff between performance and computation flexibility.

Motivation & Objective

• To address the challenge of real-time optimal passing order coordination at unsignalized intersections for connected and automated vehicles (CAVs).
• To overcome the computational infeasibility of enumerating all possible passing orders as the number of vehicles increases.
• To develop a cooperative driving strategy that balances high performance with low computation time, enabling practical deployment.
• To improve traffic efficiency and safety by minimizing total vehicle delay through intelligent scheduling.

Proposed method

• The problem is formulated as a tree search where each node represents a partial passing order, and the goal is to find the leaf node with minimum total delay.
• Monte Carlo Tree Search (MCTS) is used to explore promising branches of the solution space efficiently, prioritizing high-reward paths.
• Heuristic rules are integrated into the simulation phase to guide sampling and reduce reliance on random exploration, improving convergence speed.
• The algorithm uses a time budget of 0.1 seconds (equivalent to ~1000 nodes) to balance computation time and solution quality.
• The UCB1 formula is applied to balance exploration and exploitation during tree expansion, favoring nodes with high potential for low delay.
• A hybrid approach combines MCTS with heuristic-based simulation to prune unpromising branches early, reducing the number of explored leaf nodes from over 10^46 to around 2,000.

Experimental results

Research questions

• RQ1 Can MCTS with heuristic rules find a nearly optimal passing order for CAVs at unsignalized intersections within a sub-0.1 second planning time?
• RQ2 How does the proposed MCTS-based strategy compare in performance and computation time to FIFO and full enumeration strategies?
• RQ3 To what extent do heuristic rules improve the convergence and efficiency of MCTS in this specific tree-structured solution space?
• RQ4 What is the optimal balance between exploration and exploitation in MCTS for minimizing total vehicle delay in intersection coordination?
• RQ5 How does the performance of the proposed method scale under varying traffic densities and vehicle counts?

Key findings

• The proposed MCTS-based strategy achieved an average delay of 0.4499 seconds at 150 vehicles per hour, representing an 80% reduction compared to the FIFO strategy’s 1.3053 seconds.
• At 300 vehicles per hour, the MCTS strategy reduced average delay from 39.8313 seconds (FIFO) to 1.1407 seconds, while increasing throughput from 1,095 to 1,168 vehicles.
• The MCTS-based solution ranked 648th out of nearly 10 billion possible passing orders, indicating near-global optimality, whereas the FIFO strategy ranked 4,563,421,793rd.
• With only 1,000 nodes explored (within 0.1 seconds), the improvement rate saturated, confirming that 0.1 seconds is sufficient for high-quality solutions.
• The integration of heuristic rules reduced the number of explored leaf nodes from over 10^46 to approximately 2,000, significantly improving computational efficiency.
• The method maintained consistent performance across varying traffic demands, demonstrating robustness and scalability in diverse intersection scenarios.