A contextual framework for learning routing experiences in last-mile delivery
提出一个情境框架，通过描述性分析和预测性分析学习历史高质量路径的特征，生成类似路径，并在亚马逊最后一英里路径挑战中验证有效性，最佳算法在测试集上提升23.1%性能。
This paper presents a contextual framework for solving the experience-driven traveling salesman problem in last-mile delivery. The objective of the framework is to generate routes similar to historic high-quality ones as classified by operational experts by considering the unstructured and complex features of the last-mile delivery operations. The framework involves learning a transition weight matrix and using it in a TSP solver to generate high quality routes. In order to learn this matrix, we use descriptive analytics to extract and select important features of the high-quality routes from the data. We present a rule-based method using such extracted features. We then introduce a factorization of the transition weight matrix by features, which reduces the dimensions of the information to be learned. In the predictive analytics stage, we develop (1) Score Guided Coordinate Search as a derivative-free optimization algorithm , and (2) label-guided methods inspired by supervised learning algorithms for learning the routing preferences from the data. Any hidden preferences that are not obtained in the descriptive analytics are captured at this stage. Our approach allows us to blend the advantages of different facets of data science in a single collaborative framework, which is effective in generating high-quality solutions for a last-mile delivery problem. We test the efficiency of the methods using a case study based on Amazon Last-Mile Routing Challenge organized in 2021. A preliminary version of our rule-based method received the third place and a $25,000 award in the challenge. In this paper, we improve the learning performance of our previous methods through predictive analytics, while ensuring that the methods are effective, interpretable and flexible. Our best performing algorithm improves the performance of our rule-based method on an out-of-sample testing dataset by more than 23.1%.