基于Q学习算法的集装箱堆场翻箱落位优选1

提箱过程中的翻箱落位优化可有效降低二次及二次以上翻箱率。

箱率影响程度，选择关键因素对 Q 学习算法的多维状态空间进行表述，实时反映系统动

作的探索策略，平衡算法的收敛性和整体最优性。

位优选问题的性能，实验结果表明：1）Q 学习算法求解结果相比 Kim 翻箱量估计公式，

优化率在 40%以上；2）相比参考算法 OH 算法和 IH 算法，Q 学习算法在求解大规模问

题时，二次翻箱量平均优化率分别在 50%和 10%以上；3）不同规模贝位单个算例的 Q

学习算法求解结果相对 OH、IH 算法改进稳定，100 个算例中最多有一个算例求解结果

transportation network, and the largest collection center in the logistics supple chain. With the

deepening of goods containerization, the container throughput has increased in the terminal, the

carried out from three cases: import container reshuffle, export container reshuffle and

reshuffle when containers are moved. Then the paper analyzes the reasons for the reshuffle

from three aspects: internal factors, external factors and other irresistible factors, propose

some methods for reducing the amount of reshuffle from the aspect of yard management.

Finally, it is determined that the research object of this paper is the import containers, and it is

preferable to study the reshuffle optimization of the import containers to reduce the reshuffle

ratio of the yard.

number of a bay, and the prerequisite is that the storage status of the containers in a bay and

the key factors are selected to describe the multidimensional state space of the Q learning

algorithm to reflect the system dynamics in real time. Determine the action set and reward and

punishment system, after a certain relocation operation is completed, through the immediate

return feedback of the pros and cons of the action. Through theoretical analysis, determining

the change trend of learning factor, discount factor and exploration factor with the number of

learning scenes, and designing the exploration strategy of action, in order to balance the

convergence and overall result optimality of the algorithm.

performance of the ε- greedy Q learning algorithm which is to solve the problem of reshuffle

optimization of the import containers. The experimental results show that: 1) compared with

result is inferior to OH; at most four examples are inferior to IH.

Learning Algorithm; Markov decision process model