@phdthesis {pub5569,
	title = {Knowledge Transfer in Dynamic Multi-objective Optimization},
	author = {Gan Ruan},
	year = {2024},
	month = {April},
	abstract = {Dynamic multi-objective optimization problems (DMOPs) are optimization problems possessing multiple conflicting objectives varying with time. This type of problem widely exists in the real world, such as controlling, resource allocation and scheduling problems. Knowledge transfer, a methodology inspired by the machine learning community, has been recently used for solving DMOPs, since it can transfer useful information from solving one problem instance to solve another related problem instance, potentially speeding up the optimization process for the new instance. Although knowledge transfer has recently achieved initial successes in tackling DMOPs, there are lots of outstanding researc questions (RQs) that have never been answered. This thesis is dedicated to contributing the community of evolutionary computation from the perspective of knowledge transfer in dynamic multi-objective optimization through proposing the following RQs and respective answers (ASs). 1. Intuitively, it is impossible that any knowledge transfer approaches can be used at any cases, since transferring knowledge from an unrelated problem would be harmful to the optimization of the target problem. Therefore, RQ1 is: When and how to transfer when solving DMOPs? Only after answering these questions can we make best use of knowledge transfer to improve the optimization effectiveness. AS1: To answer this RQ, an investigation is made into when and how to transfer knowledge when leveraging knowledge transfer to tackle DMOPs with changing position and/or shape of Pareto sets (PSs) and/or Pareto fronts (PFs). We find that knowledge transfer should be avoided on problems and cases when knowledge transfer fails. Moreover, we show that different ways of how to transfer result in different quality of transferred solutions. 2. Given that knowledge transfer contributes to improving optimization quality but introduces an overhead in the optimization process, RQ2 is proposed: how efficient is the knowledge transfer, in particular could it be so time-consuming that spending the time optimizing instead of transferring knowledge would lead to similar or better results, defeating the purpose of knowledge transfer? If transfer is too time consuming, another question is whether the efficiency of transfer could be improved while preserving solution quality. Investigating these questions could enable us to have a better and more comprehensive understanding of the efficiency and effectiveness of knowledge transfer when solving DMOPs. AS2: In order to answer RQ2, a computational study is conducted on efficiency and effectiveness of knowledge transfer in dynamic multi-objective optimization. We found that the {\textquoteleft}inner{\textquoteright} optimization method introduced by transfer learning is very time-consuming. Then, we propose to use two alternatives to replace the existing one, which achieves better transfer efficiency and solutions quality. 3. Different from most other DMOPs where dynamic changes lead to changing position and/or shape of PSs and/or PFs, DMOPs with a changing number of objectives usually result in expansion or contraction of the PF or PS manifold when the number of objectives increases or reduces, respectively. Therefore, to improve the optimization process when there is a change in the number of objectives, it is intuitive to determine the expansion and contraction directions. These directions could then be used to generate transferred solutions for the new environment and further expand or contract the PSs. Therefore, RQ3 is: How to leverage knowledge transfer to heuristically determine PS expansion and contraction directions, so as to expand and contract PSs? AS3: To answer this RQ, we propose a novel knowledge transfer algorithm via PS expansion and contraction for solving DMOPs with a changing number of objectives after heuristically determining expansion and contraction directions. The algorithm is able to achieve transferred solutions with good diversity after changes, improving optimization especially in fast changing environments. 4. Moreover, it may also be possible to learn the expansion and contraction directions, so that determining them does not rely entirely on heuristics, potentially leading to more effective transfer of knowledge. Therefore, RQ4 is: How to learn the PS expansion and contraction directions when solving DMOPs with a changing number of objectives, so as to expand and contract PSs? AS4: To answer this RQ, a novel learning method for determining the expansion and contraction directions is proposed based on principal component analysis, further expanding and contracting the PSs. This approach is able to improve solution quality, not only right after changes but also after optimization of different generations.},
	publisher = {University of Birmingham},
	booktitle = {University of Birmingham}
}