Evolutionary Multiobjective Optimization Driven by Generative Adversarial Networks (GANs).

Area of Science:

• Artificial Intelligence
• Machine Learning
• Optimization Algorithms

Background:

• Evolutionary algorithms (EAs) are increasingly integrated with machine learning (ML) models.
• Model performance in ML-driven EAs depends heavily on training data quality.
• High-dimensional problems cause performance degradation due to the curse of dimensionality and limited training data.

Purpose of the Study:

• To propose a novel multiobjective evolutionary algorithm driven by generative adversarial networks (GANs).
• To address the challenge of performance deterioration in high-dimensional problems for ML-driven EAs.
• To leverage GANs' generative power for efficient solution generation with limited data.

Main Methods:

• A multiobjective evolutionary algorithm framework is developed.
• Generative adversarial networks (GANs) are employed to generate offspring solutions.
• Parent solutions are classified as real/fake samples to train GANs at each generation.

Main Results:

• The proposed GAN-driven EA generates promising offspring solutions in high-dimensional spaces.
• Effective performance is achieved even with limited training data.
• The algorithm demonstrated effectiveness on ten benchmark problems with up to 200 decision variables.

Conclusions:

• The proposed GAN-driven evolutionary algorithm effectively overcomes the curse of dimensionality.
• It offers a robust approach for generating high-quality solutions in complex, high-dimensional optimization problems.
• The method shows significant promise for advancing ML-driven evolutionary computation.