ABSTRACT

During the shipbuilding process, a block assembly line suffers a bottleneck when the largest amount of material is processed. Therefore, scheduling optimization is important for the productivity. Currently, sequence of inbound products is controlled by determining the input sequence using a heuristic or metaheuristic approach. However, the metaheuristic algorithm has limitations in that the computation time increases exponentially as the number of input objects increases, and separate optimization calculations are required for every problem. Also, the heuristic such as dispatching algorithm has the limitation of the exploring the problem domain. Therefore, this study tries a reinforcement learning algorithm based on a pointer network to overcome these limitations. Reinforcement learning with pointer network is found to be suitable for permutation flowshop problem, including input-order optimization. A trained neural network is applied without re-learning, even if the number of inputs is changed. The trained model shows the meaningful results compared with the heuristic and metaheuristic algorithms in makespan and computation time. The trained model outperforms the heuristic and metaheuristic algorithms within a limited range of permutation flowshop problem.

中文翻译：

---

使用指针网络最小化置换流水线的制造时间

摘要

在造船过程中，块装配线在处理最大量的材料时会遇到瓶颈。因此，调度优化对生产力很重要。目前，入站产品的序列是通过使用启发式或元启发式方法确定输入序列来控制的。然而，元启发式算法的局限性在于计算时间随着输入对象数量的增加呈指数增长，并且每个问题都需要单独的优化计算。此外，调度算法等启发式算法也存在探索问题域的局限性。因此，本研究尝试了一种基于指针网络的强化学习算法来克服这些限制。发现指针网络的强化学习适用于置换流水线问题，包括输入顺序优化。即使输入的数量发生变化，也无需重新学习即可应用经过训练的神经网络。与启发式和元启发式算法相比，经过训练的模型在制造时间和计算时间方面显示出有意义的结果。经过训练的模型在有限的置换流水车间问题范围内优于启发式和元启发式算法。