本文为荷兰埃因霍温理工大学（作者：A.J.G. (Arno) van de Ven）的硕士论文，共74页。

目前，NS正在开发一种局部搜索的启发式算法，它同时评估列车单元调度问题的所有组成部分，以创建调度计划。启发式算法需要从一个初始解开始。之前在NS的一项研究通过在应用启发式之前预测初始解的可行性，对这种启发式设计做出了贡献。这种方法依赖于重特征工程。本研究计划在不需要重特征工程的情况下，修正一种基于机器学习的图分类方法来执行相同的作业。结果表明，两种方法在分类精度上具有可比性。图分类方法还可以很准确地预测局部搜索启发式算法对搜索算子的评价顺序。从理论上准确预测评估顺序，可以减少搜索过程中的随机性，更快地找到改进点。虽然实际效果无法测试，但借助于图分类方法指导局部搜索启发式算法似乎是很有前途的。

A local search heuristic, currently underdevelopment at NS, evaluates all components of the Train Unit Shunting Problemsimultaneously to create shunt plans. The heuristic requires an initialsolution to start with. A previous study at NS contributed to this heuristic bypredicting feasibility of an initial solution before applying the heuristic.This method relied on heavy feature engineering. This research project modifieda machine learning based graph classification method to perform the same taskwithout the need of heavy feature engineering. Results show that both methodsare comparable in terms of classification accuracy. The graph classificationmethod is also quite accurate predicting the order in which the local searchheuristic should evaluate search operators. Accurately predicting theevaluation order theoretically leads to having less randomness in the searchprocess and finding improvements faster. Although real effects could not beentested, it seems promising to guide the local search heuristic with the help ofa graph classification method.

1. 引言

2. 项目背景与文献回顾

3. 数据产生与准备

4. DGCNN建模与结果

5. 集成学习

6. 引导式本地搜索

7. 结论与建议

更多精彩文章请关注公众号：