Investigation into Multi-objective Perishable Inventory Routing Problem with Demand Uncertainty

Abstract

The thesis aims to develop tools and provide managerial insights for perishable inventory routing problems (PIRPs), which often arise in hospitals, supermarkets, pharmaceutical companies, etc. Traditionally, decisions are made to distribute perishable items from a warehouse to receiving points while minimizing transportation and inventory holding costs. The three papers that constitute the thesis study PIRPs with increasing details, escalating uncertainty, and different scopes. In the first paper, PIRPs are studied with infinite periods, static demand, and single vehicle. The main challenge is computational due to the large variety of products and the long planning horizon. Analytical results and extensive experiments demonstrate that multi-product problems can be simplified to single-product problems under conditions of static demand, single vehicle, and a minimization objective of transportation, inventory holding, and waste costs. Further analysis reveals that optimal solutions stabilize in the long term. Based on the analysis and findings, a heuristic is developed that can provide optimal or close-to-optimal solutions for PIRPs with infinite periods and single vehicle that consumes far less computational time compared with Python+Gurobi. In the second paper, PIRPs are studied considering service level and discrete random distributions of demand with no assumptions made as to the shapes of distributions. The problems are formulated as chance-constraint models with knowledge of the convolution theorem. Equivalent deterministic approximations are derived and solved using Python+Gurobi. One could argue that the demand described by continuous distributions may be inaccurate and wrong for the following two reasons. Firstly, many perishable items are stored in discrete units such as bottles, cans, and bags. Secondly, demand can never be negative while continuous distribution functions assume that demand can take any value from negative to positive infinity. Analysis of optimal solutions shows that shapes of demand distributions significantly affect the total waste. In more detail, the more centralized and sharper the shapes of demand distributions, the less likely the waste tends to occur. The third paper studies PIRPs with multiple objectives and discrete random demand distributions. Decision makers (DMs) face difficulty picking the most appropriate solution from the pool of Pareto-optimal solutions. A decision support system (DSS) is developed to initially capture DMs' preferences by training a machine learning (ML) model based on a multi-layered belief-rule-based system (MLBRBS), then search for the most-preferred solution based on the trained MLBRBS. In this manner, instead of engaging in a lengthy and overwhelming interaction process, the DSS provides a more efficient and accurate experience. This work contributes to the literature in three ways: 1. It is the first multi-objective optimisation (MOO) method that explains its logic by an interpretable ML method with no assumptions on the shapes of utility functions; 2. A detecting mechanism is embedded to identify and report inconsistencies in DMs' preference information, which tend to accumulate as the number of objectives and the size of the Training Dataset grow; and 3. The function to provide detailed utility functions allows for more possibility to analyze DMs' behavioral logic.

Date of Award	7 May 2024
Original language	English
Awarding Institution	The University of Manchester
Supervisor	Jian-Bo Yang (Main Supervisor) & Dong Xu (Co Supervisor)