Committee-Based Active Learning for Surrogate-Assisted Particle Swarm Optimization of Expensive Problems

Area of Science:

• Computational intelligence
• Optimization algorithms
• Machine learning applications

Background:

• Expensive function evaluations (FEs) challenge evolutionary algorithms (EAs) in real-world optimization.
• Existing surrogate-assisted EAs (SAEAs) often require numerous FEs or are limited to low-dimensional problems.
• Need for efficient SAEAs that balance accuracy and computational cost.

Purpose of the Study:

• To propose a novel surrogate-assisted particle swarm optimization (PSO) algorithm.
• To address the challenge of expensive function evaluations in optimization problems.
• To improve the efficiency and applicability of SAEAs to higher-dimensional problems.

Main Methods:

• Developed a global model management strategy inspired by committee-based active learning (CAL).
• Utilized a surrogate ensemble to identify promising and uncertain solutions for evaluation.
• Integrated a local surrogate model for focused optimization around the best-found solution.
• Implemented a switching mechanism between global and local search strategies based on performance.

Main Results:

• The proposed algorithm achieved better or competitive solutions compared to state-of-the-art SAEAs.
• Demonstrated effectiveness on benchmark problems with up to 30 decision variables.
• Successfully applied to a practical airfoil design optimization problem.
• Required a significantly limited budget of hundreds of exact FEs.

Conclusions:

• The novel CAL-inspired SPSO algorithm offers an efficient approach to computationally expensive optimization problems.
• The hybrid global-local search strategy effectively reduces the number of required function evaluations.
• The method shows promise for both theoretical benchmarks and real-world engineering applications.