Machine-Learning-Based Prediction of MultiCompartment Vehicle Fleet P...

We present a method to perform a comprehensive analysis of the fleet composition problem that is suitable for most variants of the vehicle routing problem. Its basic principle is to estimate a fleet’s performance by using the company’s delivery planning tools in a black-box fashion. In a case study, we analyze the fleet size and mix for a fictional grocery home delivery service. A fleet comprising multi-compartment vehicles is employed, where each compartment is designated for storing groceries at specific temperature zones tailored to their storage requirements. In general, the stakeholders are interested in finding a fleet configuration that enables good performance regarding defined key performance indicators (KPIs). Seasonal demand changes occur in nearly all types of routing applications. Therefore, we aim to identify fleet configurations that ensure consistent and satisfactory performance across all seasons. We do not propose a methodology for choosing a fleet. This is because stakeholders may consider multiple KPIs when making fleet composition decisions, and these KPIs may be conflicting and vary by scenario. Thus, we focus on a method for predicting the values of multiple KPIs for a given fleet.