Automated design of local search algorithms for vehicle routing problems with time windows

Meng, Weiyao (2023) Automated design of local search algorithms for vehicle routing problems with time windows. PhD thesis, University of Nottingham.

PDF (Thesis - as examined) - Requires a PDF viewer such as GSview, Xpdf or Adobe Acrobat Reader
Available under Licence Creative Commons Attribution.
Download (3MB) | Preview


Designing effective search algorithms for solving combinatorial optimisation problems presents a challenge for researchers due to the time-consuming experiments and experience required in decision-making. Automated algorithm design removes the heavy reliance on human experts and allows the exploration of new algorithm designs. This thesis systematically investigates machine learning for the automated design of new and generic local search algorithms, taking the vehicle routing problem with time windows as the testbed.

The research starts by building AutoGCOP, a new general framework for the automated design of local search algorithms to optimise the composition of basic algorithmic components. Within the consistent AutoGCOP framework, the basic algorithmic components show satisfying performance for solving the VRPTW. Based on AutoGCOP, the thesis investigates the use of machine learning for automated algorithm composition by modelling the algorithm design task as different machine learning tasks, thus investigating different perspectives of learning in automated algorithm design.

Based on AutoGCOP, the thesis first investigates online learning in automated algorithm design. Two learning models based on reinforcement learning and Markov chain are investigated to learn and enhance the compositions of algorithmic components towards automated algorithm design. The Markov chain model presents a superior performance in learning the compositions of algorithmic components during the search, demonstrating its effectiveness in designing new algorithms automatically.

The thesis then investigates offline learning to learn the hidden knowledge of effective algorithmic compositions within AutoGCOP for automated algorithm design. The forecast of algorithmic components in the automated composition is defined as a sequence classification task. This new machine learning task is then solved by a Long Short-term Memory (LSTM) neural network which outperforms various conventional classifiers. Further analysis reveals that a Transformer network surpasses LSTM at learning from longer algorithmic compositions. The systematical analysis of algorithmic compositions reveals some key features for improving the prediction.

To discover valuable knowledge in algorithm designs, the thesis applies sequential rule mining to effective algorithmic compositions collected based on AutoGCOP. Sequential rules of composing basic components are extracted and further analysed, presenting a superior performance of automatically composed local search algorithms for solving VRPTW. The extracted sequential rules also suggest the importance of considering the impact of algorithmic components on optimisation performance during automated composition, which provides new insights into algorithm design.

The thesis gains valuable insights from various learning perspectives, enhancing the understanding towards automated algorithm design. Some directions for future work are present.

Item Type: Thesis (University of Nottingham only) (PhD)
Supervisors: Qu, Rong
Landa-Silva, Dario
Keywords: Automated Algorithm Design, Vehicle Routing, Combinatorial Optimisation, Machine Learning, Data Mining
Subjects: Q Science > QA Mathematics > QA 75 Electronic computers. Computer science
Faculties/Schools: UK Campuses > Faculty of Science > School of Computer Science
Item ID: 76313
Depositing User: Meng, Weiyao
Date Deposited: 12 Dec 2023 04:40
Last Modified: 12 Dec 2023 04:40

Actions (Archive Staff Only)

Edit View Edit View