Achieving Highly Scalable Evolutionary Real-Valued Optimization by Exploiting Partial Evaluations.

Area of Science:

• Computational intelligence
• Optimization algorithms
• Evolutionary computation

Background:

• Efficient scalability in Evolutionary Algorithms (EAs) requires accounting for dependencies (linkage) during variation.
• Gray-Box Optimization (GBO) settings benefit from exploiting prior knowledge of these dependencies.
• Partial evaluations, where modified solutions are efficiently assessed, are crucial for difficult problems like non-separable, multimodal, and multiobjective optimization.

Purpose of the Study:

• To introduce the Real-Valued GOMEA (RV-GOMEA) for real-valued optimization problems.
• To present a new RV-GOMEA variant combining GOMEA with Covariance Matrix Adaptation Evolution Strategies (CMA-ES).
• To evaluate the performance and scalability of RV-GOMEA variants in a GBO setting.

Main Methods:

• Developed Real-Valued GOMEA (RV-GOMEA) by extending the Gene-pool Optimal Mixing Evolutionary Algorithm (GOMEA) for real-valued domains.
• Created a novel RV-GOMEA variant by integrating GOMEA with Covariance Matrix Adaptation Evolution Strategies (CMA-ES).
• Compared RV-GOMEA variants against L-BFGS and Limited Memory CMA-ES (LM-CMA-ES) in a GBO context.

Main Results:

• Both RV-GOMEA variants demonstrated excellent performance and scalability in the GBO setting.
• The proposed RV-GOMEA approaches achieved performance orders of magnitude better than EAs not exploiting the GBO setting.
• Effective exploitation of partial evaluations by RV-GOMEA leads to substantial improvements.

Conclusions:

• RV-GOMEA effectively leverages partial evaluations in GBO for improved optimization.
• The developed RV-GOMEA variants offer superior performance and scalability for challenging optimization tasks.
• This work establishes RV-GOMEA as a powerful tool for real-valued GBO.