OPT-GAN: A Broad-Spectrum Global Optimizer for Black-Box Problems by Learning Distribution

Black-box optimization (BBO) algorithms are concerned with finding the best solutions for problems missing analytical details. Most classical methods such based on strong and fixed a priori assumptions, as Gaussianity. However, complex real-world problems, especially when global optimum is desired, could be very far from assumptions because of their diversities, causing unexpected obstacles. In this study, we propose generative adversarial net-based broad-spectrum optimizer (OPT-GAN) which estimates distribution gradually, strategies to balance exploration-exploitation trade-off. It has potential better adapt regularity structure diversified landscapes than other prior, e.g., Gaussian assumption or separability. Experiments diverse BBO benchmarks high dimensional real world applications exhibit that OPT-GAN outperforms traditional neural algorithms. The code Appendix available at https://github.com/NBICLAB/OPT-GAN