Date of Award
Master of Science in Systems Engineering
Mechanical, Industrial and Systems Engineering
The Vehicle Routing Problem (VRP) is a widely known NP-Hard operations research problem endowed with a wide range of heuristic algorithms generated from decades of global research. These heuristics provide solutions within a reasonable run time, but at some expense to optimality. The literature further suggests heuristic performance in one class of problems comes at a cost in performance to other classes. This study aimed to develop methods for selecting the best heuristic from a defined set to solve an arbitrary VRP instance.
Known as the algorithm selection problem, this study implemented supervised machine learning techniques to construct prediction models based upon instance characteristics. These models were evaluated by metrics commonly found in both algorithm selection and machine learning studies. Built from a set of 23 features and a portfolio of four varied heuristics, the leading model correctly predicted the best algorithm with 79% accuracy despite the single best heuristic occurring only 49% of the time.
Models were constructed using a custom problem space of 5,000 VRP instances developed organically by novel methods adapted from the literature. Adequacy of the problem space, regarding its range of difficulties and sufficiency of size, was also explored. The results indicate the problem space was appropriately diverse and the prediction models, which were developed by learning algorithms using provided data, are unlikely to improve accuracy if given more data.
Fellers, Justin C., "ALGORITHM SELECTION FOR THE CAPACITATED VEHICLE ROUTING PROBLEM USING MACHINE LEARNING CLASSIFIERS" (2021). Open Access Master's Theses. Paper 1933.