This PhD thesis addresses the challenge of developing a tool to provide algorithmic recommendation to end users (experts in the problem domain but not experts in multi-objective algorithms) without the need of a resource-intensive process of auto-configuration. This challenge is faced with an approach based on previous knowledge about the problems.
A semantic model, moody, is designed to formally define knowledge in the field of multi-objective optimization with metaheuristics, with a focus on the relevant concepts required to characterize problems and the performance of algorithms.
moorphology is developed as a tool to provide landscape characteristics of the search and objective spaces of multi-objective problems. These landscape characteristics are a key factor for the computation of a similarity metric between multi-objective problems, which are a necessity to provide recommendations based on previous knowledge.
To generate in an efficient way the required knowledge to implement the recommendation engine, a meta-optimization approach is presented as the software tool Evolver. This tool allows the automatic configuration of metaheuristics by defining it as an optimization problem.
Large language models are evaluated for the task of helping domain experts in implementing their problems into an optimization framework for solving them. To solve this problem, a large language model is fine-tuned and embedded into a graphical tool, named moostral, to allow the end user to easily implement their optimization framework into the recommendation system described in this thesis.
To connect the previously mentioned elements, a recommendation engine, named recommoonder, is implemented to solve the challenge presented in this thesis.
This thesis has a very practical focus, providing open source repositories for all the tools developed in it, allowing their use in the further research lines defined in the last chapter.
Buscar en Dimension