(Kirley, Michael)(Munoz Acosta, Mario)Alsouly, Hanan2025-08-042025https://hdl.handle.net/20.500.14154/76092Constrained multi-objective optimization problems (CMOPs) present significant challenges due to the need to optimize multiple conflicting objectives simultaneously under various constraints. In recent years, several constrained multi-objective evolutionary algorithms (CMOEAs) have been developed to tackle these problems. However, their performance is typically evaluated by comparing a few algorithms on specific problem instances, which can make it difficult to know which algorithm is the best fit for a particular situation. Selecting the most suitable algorithm for a given problem instance, known as algorithm selection, is critical yet challenging due to the diverse and complex nature of CMOPs. Automating the algorithm selection process can be accomplished by delegating the task to a machine learning model. These models often require information about the problem to make reasonable decisions. A prominent approach to provide such information is landscape analysis, where the characteristics of a problem’s landscape are captured through numerical features. Despite the importance of algorithm selection, a notable gap remains in applying it to the CMOP domain, primarily due to the lack of sufficient landscape features and studies that explore how CMOP characteristics relate to CMOEA performance. This thesis addresses the existing gap by introducing innovative landscape analysis features specifically designed for CMOPs. These features capture essential elements of a problem’s landscape, such as the relationship between constraint violations and multi-objective fitness. Through instance space analysis, we demonstrate how these features highlight the strengths and weaknesses of various CMOEAs. Building on these insights, we propose a framework that incorporates algorithm selection directly into the optimization process. By combining our new landscape features, the performance data learned from benchmarking CMOEAs, and various machine learning methods, our framework enables the on-the-fly selection of the best-performing algorithm for each problem instance. Our research demonstrates that incorporating landscape features into the optimization process can lead to improved algorithm performance. Through the novel landscape features and the development of practical algorithm selection models, this research significantly advances the field of constrained multi-objective optimization.177enoptimisationalgorithm selectionlandscape analysisevolutionary algorithmevolutionary computationconstrained multi-objective optimisationinstance space analysisThesis