Titulo Estágio
Performance Evaluation of Multiobjective Optimization Algorithms
Áreas de especialidade
Sistemas Inteligentes
Local do Estágio
DEI
Enquadramento
Real-world problems typically involve multiple incommensurable objectives such as cost, performance, and environmental impact. Solving such problems amounts to finding solutions that perform sufficiently well according to all objectives for practical purposes and cannot be improved further in all objectives simultaneously. In practice, there is usually no perfect, optimal solution but rather a set of solutions termed efficient in the sense that they reflect the trade-offs between the objectives of the problem.
The last 30 years have seen a growing interest in the development of algorithms for multiobjective optimization, especially for the case where the mathematical structure of the objective functions is not directly available to the algorithm or known at all – the so-called black-box case. Approaches to this kind of problem include mathematical optimization and meta-heuristic algorithms such as direct-search and evolutionary algorithms, respectively.
While the merits of the different algorithms are inevitably touted and typically demonstrated by their authors through experimental studies, it is usually less clear what kind of problems may make those algorithms fail. Independent benchmarking studies and competitions do try to address this issue to some extent, but there is still room for improvement with respect to the experimental methodology adopted, as the recent interest of the community in optimization algorithm benchmarking suggests.
Objetivo
The main objective of this work is to perform a well-grounded experimental evaluation of the performance of state-of-the-art algorithms for multiobjective optimization in continuous domains. The study will cover the selection of benchmark problem instances, the identification of the optimization algorithms of interest, the definition of suitable performance measures, the design and conduction of experiments, data analysis and visualization, and reporting. Aspects of interest include, but are not limited to, algorithm scalability with respect to the number of decision variables and/or objective functions, possible interactions between problems and algorithms, and the extent to which the conclusions drawn generalize beyond the study itself.
Plano de Trabalhos - Semestre 1
1. Literature review on multiobjective optimization algorithms and benchmarking
2. Study of existing multiobjective optimization benchmark suites and performance indicators
3. Experimental setup, preliminary experiments, and exploratory data analysis
4. Tentative formulation of research questions for further study
5. Intermediate report writing
Plano de Trabalhos - Semestre 2
1. Experiment design
2. Conduction of experiments
3. Data analysis and reporting
4. External validity assessment
5. Dissertation writing
Condições
Good background in mathematical and/or heuristic optimization, statistics, and good programming skills. The work will be carried out in the laboratories of the CISUC Adaptive Computing Group. A multiprocessor server is available to run experiments and analyse data.
A full-time research scholarship (Bolsa de Investigação para Licenciado) to support this line of work is planned for the second semester.
Orientador
Carlos Manuel Mira da Fonseca
cmfonsec@dei.uc.pt 📩