Data-Driven Evolutionary Algorithm With Perturbation-Based Ensemble Surrogates.

Area of Science:

• Computational Intelligence
• Optimization Algorithms
• Machine Learning

Background:

• Data-driven evolutionary algorithms (DDEAs) leverage data and surrogate models for efficient optimization, especially for expensive objective functions.
• DDEA performance is sensitive to surrogate quality and data availability, often degrading with less data.

Purpose of the Study:

• To propose a new DDEA framework, DDEA-PES, designed to improve surrogate accuracy and robustness, particularly in low-data scenarios.
• To enhance data utilization and overall optimization efficiency in DDEAs.

Main Methods:

• Development of a DDEA framework incorporating perturbation-based ensemble surrogates (DDEA-PES).
• Implementation of two key mechanisms: diverse surrogate generation via data perturbation and selective ensemble modeling.
• Validation through theoretical analysis and experimental comparisons using genetic algorithms and radial basis function networks.

Main Results:

• The proposed DDEA-PES framework demonstrates advantages in data quantity, utilization, and surrogate accuracy.
• Experimental results on benchmarks and an aerodynamic airfoil design problem show DDEA-PES outperforms state-of-the-art DDEAs.
• DDEA-PES achieved competitive results using only approximately 2% of the computational budget compared to traditional non-data-driven methods.

Conclusions:

• The DDEA-PES framework offers a robust and efficient approach to optimization problems with expensive objective functions.
• The perturbation-based ensemble surrogate strategy significantly improves DDEA performance and reduces computational requirements.
• DDEA-PES presents a promising advancement for practical applications requiring high-accuracy optimization with limited data.