Author(s)

WenJie Wu

Affiliation(s)

¹Department of Vehicle Engineering, Tongji University, Shanghai 200092, China.

²Department of Supplier Technical Assistance, Jiangling Motors Co., Ltd., Nanchang 330000, Jiangxi, China.

Corresponding Author

WenJie Wu

ABSTRACT

With the continuous advancement of autonomous driving technology in complex urban traffic environments, achieving efficient and safe trajectory planning in complex intersection scenarios with frequent vehicle interactions and dynamic obstacles has become one of the core challenges in current research. As an important structure for representing multimodal driving behaviors, the trajectory tree plays a key role in the decision-making process at complex intersections. However, its large search space and high computational complexity severely limit real-time performance and scalability. To address this, this paper focuses on the optimization problem of trajectory trees in complex intersections and innovatively introduces reinforcement learning algorithms. A spatiotemporal joint pruning strategy based on reinforcement learning is proposed to improve the search efficiency and decision-making quality of the trajectory tree. This strategy effectively reduces redundant trajectory branches by combining spatial and temporal pruning mechanisms, dynamically adjusting the search direction, and thus achieves precise control and efficient evolution of the trajectory tree. In terms of model design, this paper systematically defines the action set, state space, and reward function, ensuring that the reinforcement learning agent can learn pruning strategies with generalization capability in complex traffic environments. Furthermore, the paper improves the original model, clarifying the trajectory tree optimization goals and enhancing the adaptability and stability of the strategy. In the experimental section, representative urban traffic datasets are selected, with reasonable parameter configurations and evaluation metrics. The proposed method is comprehensively evaluated from three dimensions: trajectory clustering performance, path optimization, and computational efficiency, and compared with several mainstream methods. Experimental results demonstrate that the proposed spatiotemporal joint pruning strategy based on reinforcement learning exhibits significant effectiveness and superiority in trajectory tree optimization for complex intersections. It not only improves the vehicle’s passing ability in dynamic environments but also provides reliable technical support for the deployment of autonomous driving systems in real-world scenarios, with important theoretical value and engineering application prospects.

KEYWORDS

Reinforcement learning; Spatiotemporal joint pruning strategy; Trajectory tree optimization; Complex intersection

CITE THIS PAPER

WenJie Wu. Spatiotemporal joint pruning strategy based on reinforcement learning for trajectory tree optimization in complex intersection applications. Journal of Computer Science and Electrical Engineering. 2025, 7(6): 1-7. DOI: https://doi.org/10.61784/jcsee3081.

REFERENCES

[1] Xu Y, Wu C. Path planning for vertical parking in complex environments based on an improved Rapidly-exploring Random Tree. Automobile Technology, 2022(07): 42-47. DOI:10.19620/j.cnki.1000-3703.20210717.

[2] Lu Y, Kuang J, Xiao Y, et al. Heuristic search-based multi-agent asymmetric trajectory planning. Computer Engineering and Applications, 2025, 61(02): 344-354.

[3] Wang Z. Path planning and trajectory optimization based on the improved A-star algorithm. Jilin University, 2024. DOI:10.27162/d.cnki.gjlin.2024.004857.

[4] Wang X, Liu P, Xiang S, et al. Unstructured pruning methods based on neural architecture search. Pattern Recognition and Artificial Intelligence, 2023(05): 448-458. DOI:10.16451/j.cnki.issn1003-6059.202305005.

[5] Ruan X, Hu W, Liu Y, et al. Model pruning based on dynamic sparsity and feature learning enhancement. Science China: Technological Sciences, 2022(05): 667-681.

[6] Wan Z, Li L, Yang M, et al. Decision tree model for extracting road intersection features from vehicle trajectory data. Journal of Surveying and Mapping, 2019(11): 1391-1403.

[7] Weng J, Qin Y, Tang X, et al. Lane detection model pruning algorithm based on multi-objective optimization. Sensors and Microsystems, 2023(07): 125-127+131. DOI:10.13873/J.1000-9787(2023)07-0125-03.

[8] Liu Y, Teng Y, Niu T, et al. Filter pruning scheme based on deep reinforcement learning. Journal of Beijing University of Posts and Telecommunications, 2023(03): 31-36. DOI:10.13190/j.jbupt.2022-140.

[9] Gomes R, Kowalczyk R. Dynamic analysis of multiagent Q-learning with ε-greedy exploration, 2009.

[10] Yang J. Research on multi-layer composite pruning and its acceleration algorithm based on convolutional neural networks. Changsha University of Science & Technology, 2021. DOI:10.26985/d.cnki.gcsjc.2021.000196.

[11] Krishnaswamy R, Li S, Sandeep S. Constant approximation for k-median and k-means with outliers via iterative rounding, 2018.

[12] Guo C. Hybrid pruning research based on convolutional neural networks. Changsha University of Science & Technology, 2021. DOI:10.26985/d.cnki.gcsjc.2021.000279.

[13] Yi X, Ma S, Xiao N. Run-time optimization of deep learning accelerators under different pruning strategies. Computer Engineering and Science, 2023(07): 1141-1148.

[14] Qu H, Zhang X, Wang Y. Pruning method for convolutional neural networks based on information fusion strategies. Computer Engineering and Applications, 2022(24): 125-133.