Overview
This project benchmarks ten multi-objective optimization algorithms, including NSGA-II, NSGA-III, SPEA2, MOEA/D, HYPE, NSPSO, MACO, RBFMOpt, TPE, and NOMAD, across a large and diverse set of black-box (bbcom) problems. The study evaluates algorithm performance, robustness, convergence, and sensitivity to hyperparameters.
Hyperparameter Tuning Approach
A three-stage hyperparameter tuning approach was applied:
- Random Sampling: Hyperparameter configurations were randomly sampled and evaluated on 40 problems, resulting in 2,000 runs.
- Random Forest Sensitivity Analysis: Random Forest importance analysis was used to identify the most influential hyperparameters for each algorithm.
- Surrogate-Based Optimization: RBFOpt was used to optimize the hyperparameters according to four scenarios:
- Optimizing all parameters.
- Optimizing the four most important parameters.
- Optimizing all parameters with a threshold of 0.15.
- Optimizing parameters with a total importance sum of 0.8.
Evaluation
The tuned algorithms were subsequently compared with their default configurations across 3,960 problems to assess the performance improvements resulting from hyperparameter optimization.
Objective
Overall, the project aims to identify which multi-objective optimization algorithms and hyperparameter configurations provide reliable and effective performance across a diverse range of multi-objective black-box optimization problems.
(2026-08-12)
