A New Surrogate-Assisted Interactive Genetic Algorithm With Weighted Semisupervised Learning.

Area of Science:

• Computational Intelligence
• Optimization Algorithms
• Human-Computer Interaction

Background:

• Surrogate-assisted evolutionary algorithms (SAEAs) reduce computational cost but struggle with subjective user feedback.
• Interactive genetic algorithms (IGAs) incorporate human judgment but can lead to user fatigue.
• Existing surrogates in IGAs often fail to account for uncertainty in human-assigned fitness values.

Purpose of the Study:

• To develop a novel surrogate-assisted interactive genetic algorithm (IGA) that effectively handles uncertainty in subjective user evaluations.
• To enhance surrogate model accuracy and improve the management of surrogate models by exploiting uncertainty.
• To reduce human fatigue in complex design optimization problems.

Main Methods:

• A new surrogate-assisted IGA is proposed, incorporating uncertainty from interval-based fitness values.
• An improved cotraining algorithm for semisupervised learning is used, considering uncertainty in training and weighting models.
• Model management prioritizes re-evaluation of both high-performing and uncertain individuals.

Main Results:

• The proposed algorithm effectively alleviates user fatigue in interactive optimization.
• Exploiting uncertainty in subjective evaluations enhances surrogate approximation accuracy.
• The method demonstrates effectiveness on test problems and in a fashion design application.

Conclusions:

• The novel surrogate-assisted IGA successfully addresses the challenge of uncertainty in subjective fitness assignments.
• This approach leads to reduced user fatigue and improved identification of acceptable solutions for complex design problems.
• The findings suggest a promising direction for human-in-the-loop optimization systems.