Kernel method based human model for enhancing interactive evolutionary optimization.

Area of Science:

• Artificial Intelligence
• Evolutionary Computation
• Human-Computer Interaction

Background:

• Fitness landscapes are crucial in evolutionary computation (EC) but often lack accuracy in discrete spaces, especially for interactive EC (IEC).
• Modeling human subjective evaluation in IEC is challenging due to its complexity and difficulty in defining the fitness landscape.
• Existing EC methods struggle with the nuanced perceptual knowledge derived from human feedback.

Purpose of the Study:

• To propose a novel method for establishing a human model in a projected high-dimensional search space for enhancing IEC.
• To leverage kernel classification and bivalent logic principles to create a predictive human model.
• To develop an evolution control method integrated with the human model to improve IEC search efficiency.

Main Methods:

• Establishing a human model using kernel classification in a projected high-dimensional feature space.
• Designing a linear classifier within the feature space to capture user preference knowledge.
• Implementing an evolution control strategy based on the predictive human model to guide the search process.
• Utilizing bivalent logic as a foundational principle for modeling human perception.

Main Results:

• The proposed human model effectively predicts user preferences and perceptual knowledge.
• The integrated evolution control method significantly enhances the performance of interactive EC search.
• Experimental results with a pseudo-IEC user demonstrate substantial improvements in search efficiency.

Conclusions:

• The developed human model and evolution control method offer a significant advancement for interactive evolutionary computation.
• This approach addresses the limitations of traditional fitness landscapes in capturing human subjective evaluations.
• The method provides a robust framework for enhancing EC search through accurate modeling of human perception.