Evolution by adapting surrogates.

Area of Science:

• Computational Intelligence
• Optimization Algorithms
• Machine Learning

Background:

• Surrogate-assisted evolutionary algorithms are crucial for computationally expensive optimization problems.
• Current frameworks often use ad hoc approximation methods, lacking adaptability.
• Prior knowledge of suitable approximation methodologies is typically unavailable.

Purpose of the Study:

• To present a novel evolutionary framework, Evolvable Learning of Surrogates (EvoLS), that adaptively learns and utilizes diverse approximation methodologies.
• To introduce 'evolvability' as a criterion for selecting surrogates, focusing on fitness improvement rather than prediction error.
• To enable self-configuration of surrogate-assisted memetic algorithms for computationally expensive problems.

Main Methods:

• EvoLS operates on multiple diverse approximation methodologies within the evolutionary search.
• A statistical learning scheme determines the evolvability of each approximation methodology online.
• The most productive approximation methodology (highest evolvability) is inferred for each solution to construct fitness-improving surrogates.
• A trust-region enabled local search strategy is integrated for enhanced optimization.

Main Results:

• EvoLS demonstrated reliable, high-quality, and efficient performance on benchmark problems.
• The framework proved effective in solving a real-world computationally expensive aerodynamic car rear design problem.
• Adaptive selection of approximation methodologies led to superior optimization outcomes under limited computational budgets.

Conclusions:

• EvoLS offers an effective and self-configuring approach to surrogate-assisted evolutionary optimization.
• The concept of 'evolvability' provides a robust criterion for adaptive surrogate selection.
• The framework significantly enhances performance for computationally demanding optimization tasks.