基于Mamba-2编码的集装箱箱位分配模型  

Container Slot Assignment Model Based on Mamba-2 Encoding

作　　者：向若愚 杨有[2] 陈雁翎 XIANG Ruoyu;YANG You;CHEN Yanling(School of Computer and Information Science,Chongqing Normal University,Chongqing 401331,China;National Center for Applied Mathematics in Chongqing,Chongqing Normal University,Chongqing 401331,China)

机构地区：[1]重庆师范大学计算机与信息科学学院,重庆401331 [2]重庆师范大学重庆国家应用数学中心,重庆401331

出　　处：《河南科学》2025年第3期321-329,共9页Henan Science

基　　金：重庆市自然科学基金创新发展联合基金项目(CSTB2023NSCQ-LZX0142);重庆市教委科技重点项目(KJZDK202400503)。

摘　　要：箱位分配在集装箱码头中至关重要,影响非生产性成本和作业效率。针对箱位分配问题,基于规则的策略求解速度快,但理论上无法获取最优解;数学规划模型理论上可以获得最优解,但计算时间随堆场规模增加而呈指数型增长,难以满足集装箱堆存的应用要求。使用深度强化学习方法设计模型,可以在短时间内获得高质量解。为此,针对贝位构造不能充分表达栈间语义的问题,定义5个栈位输入特征;设计基于端到端的编解码器模型,用于集装箱箱位分配。该模型采用Mamba-2进行编码,使用多头注意力进行解码,使用带基线的强化学习算法进行训练。仿真实验表明,所设计模型在中大规模问题上具有性能优势,能在较短时间内选择合理箱位,降低翻箱率,提高港口作业效率。The assignment of container slots is crucial in container terminals,significantly impacting nonproductive costs and operational efficiency.In addressing the container slot assignment problem,rule-based strategies provide fast solution,but they are theoretically incapable of obtaining the optimal solution.While mathematical programming models can theoretically achieve optimality,their computational time increases exponentially with the scale of the yard,making them impractical for container storage applications.A deep reinforcement learning-based approach,however,is capable of obtaining high-quality solutions within a short time.To address the issue that the block configuration fails to fully express the semantic meaning between stacks,five input features of stack are defined.An end-to-end encoder-decoder model is designed for container slot assignment,by utilizing Mamba-2 for encoding,multi-head attention for decoding,and a reinforcement learning algorithm with a baseline for training.According to the simulation results,the proposed model exhibits performance advantages in medium to large-scale problems,enabling the selection of reasonable slots within a shorter time,reducing the container relocation rate,and enhancing port operational efficiency.

关 键 词：集装箱 箱位分配 Mamba-2 深度强化学习 状态空间模型 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]