Jackson, Warren G
(2021)
Move Acceptance in Local Search Metaheuristics for Cross-domain Heuristic Search.
PhD thesis, University of Nottingham.
Abstract
Many real-world combinatorial optimisation problems (COPs) are computationally hard problems and search methods are frequently preferred as solution techniques. Traditionally, an expert with domain knowledge designs, and tailors the search method for solving a particular COP.
This process is usually expensive, requiring a lot of effort and time and often results in problem specific algorithms that can not be applied to another COP. Then, the domain expert either needs to design a new search method, or reconfigure an existing search method to solve that COP.
This prompted interest into developing more general, problem-domain-independent high-level search methods that can be re-used, capable of solving not just a single problem but multiple COPs.
The cross-domain search problem is a relatively new concept and represents a high-level issue that involves designing a single solution method for solving a multitude of COPs preferably with the least or no expert intervention.
Cross-domain search methods are algorithms designed to tackle the cross-domain search problem. Such methods are of interest to researchers and practitioners worldwide as they offer a single off-the-shelf go-to approach to problem solving.
Furthermore, if a cross-domain search method has a good performance, then it can be expected to solve `any' given COP well and in a reasonable time frame.
When a practitioner is tasked with solving a new or unknown COP, they are tasked with a decision-making dilemma. This entails the decision of what algorithm they should use, what parameters should be used for that algorithm, and whether any other algorithm can outperform it.
A well designed cross-domain search method that performs well and does not require re-tuning can fulfil this dilemma allowing practitioners to find good-enough solutions to such problems.
Researchers on the other hand strive to find high-quality solutions to these problems; however, such a cross-domain search method provides them with a good benchmark to which they can compare their solution methods to, and should ultimately aim to outperform.
In this work, move acceptance methods, which are a component of traditional search methods, such as metaheuristics and hyper-heuristics, are explored under a cross-domain search framework.
A survey of the existing move acceptance methods as a part of local search metaheuristics is conducted based on the hyper-heuristic literature as solution methods to the cross-domain search problem.
Furthermore, a taxonomy is provided for classifying them based on their design characteristics.
The cross-domain performance of existing move acceptance methods, covering the taxonomy, is compared across a total of 45 problem instances spanning 9 problem domains, and the effects of parameter tuning versus choice of the move acceptance method are explored.
A novel move acceptance method (HAMSTA) is proposed to overcome the shortcomings of the existing methods to improve the cross-domain performance of a local search metaheuristic.
HAMSTA is capable of outperforming the cross-domain performances of existing methods that are re-tuned for each domain, despite itself using only a single cross-domain parameter configuration derived from tuning experiments that considers 2 instances each from 4 domains; hence, HAMSTA requires no expert intervention to re-configure it to perform well for solving multiple COPs with 37 problem instances unseen by HAMSTA, 25 of which are from unseen domains.
HAMSTA is therefore shown to have the potential to fulfil the aforementioned decision-making dilemma.
Actions (Archive Staff Only)
 |
Edit View |
Tools
Tools
