Abstract
Humans have an outstanding capability to transfer knowledge across different tasks they encounter daily. When faced a new task, we typically draw on our past experiences and apply relevant ideas or skills to solve it, rather than treating it as a brand-new challenge and starting from scratch, enabling us to tackle many tasks with ease and efficiency. Recently, this outstanding ability has motivated researchers to incorporate it into optimization algorithms towards improved search performance, giving rise to an emerging research topic known as transfer optimization. Specifically, the ultimate goal of transfer optimization is to improve the search performance of an optimization algorithm on target task(s) of interest by leveraging knowledge extracted from possibly related source tasks.In cases of frequent problem-solving for optimization tasks within a particular domain due to changing conditions or problem features, a growing number of tasks will be solved and stored in a database, providing an opportunity for the current target task to achieve better results through knowledge transfer from the previously-solved tasks, which is formally known as sequential transfer optimization. Despite the variety of sequential transfer optimization algorithms developed over the decades, an in-depth study of common benchmark problems for comprehensive evaluation, as well as well-grounded foundations for rationalizing knowledge transfer and pertinent algorithm designs with sound performance gain, remains underexplored. In response to this gap, this thesis is done in the context of continuous optimization with evolutionary computation due to its well-recognized generality and applicability, wherein the optimized solutions of the previously-solved tasks serve as knowledge to be transferred across tasks based on analogical reasoning, to contribute to the aforementioned three aspects of sequential transfer optimization, as summarized in what follows.
Firstly, a benchmark suite of sequential transfer optimization problems (STOPs) is developed for transfer algorithms whose underlying method of reasoning for transferring previously-optimized solutions is analogy: decisions about knowledge transfer between two tasks are made according to their similarities. Particularly, a problem generator with configurable source-target similarities is proposed to enable a close resemblance to real-world STOPs. Thereafter, a large body of analogy-based knowledge transfer methods are employed to empirically validate the superiority of the benchmark problems developed in this thesis over existing test problems, with the underlying rationales theoretically analyzed according to the sub-processes in analogical reasoning. In addition, two important theorems about the performance gain of transferring optimized solutions are presented, namely unconditionally non-negative performance gain (UNPG) and conditionally positive performance gain (CPPG), which serve as a beacon for guiding subsequent algorithm designs.
Secondly, to achieve the goal of UNPG, a novel competitive knowledge transfer (CKT) method is proposed, which treats both the source optimized solutions and the target solutions as task-solving knowledge from a consistent perspective, enabling them to compete with each other to elect the winner for evaluation. As a consequence, the target search is expected to be greatly improved in cases of promising previously-optimized solutions available in the database. Moreover, with a novel concept named optimization progress, the performance gain brought by CKT is analyzed analytically. Experimental studies conducted on the benchmark suite and a practical case study from the petroleum industry have validated the efficacy of CKT.
Thirdly, to approach the goal of CPPG, a dual knowledge transfer (DKT) method is proposed for problems with strong task heterogeneities in both decision and objective spaces, which employs two distinct transfer methods to leverage convergence-related and diversity-related search experiences cooperatively. Specifically, through decision variable analysis, a convergence transfer method is proposed to speed up the target search towards the optimum of convergence-related variables, while a diversity transfer method is developed to improve the distribution of diversity-related variables. Empirical results on a set of objective-heterogeneous problems and a practical application from the mineral processing have verified the effectiveness of DKT.
This thesis aims not only to develop effective algorithms for leveraging optimization experience in the form of optimized solutions from previously-solved tasks but also to convey the philosophy reflected in the research pathway paved by the common fundamentals of analogy-based knowledge transfer, which would stimulate further research on sequential transfer optimization with other novel transferable objects.
| Date of Award | 30 Aug 2024 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Linqi SONG (Supervisor) & Kay Chen Tan (External Co-Supervisor) |