Semantic variation operators for multidimensional genetic programming

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Genetic programming (GP) represents solutions as programs, offering potential for building block identification.
• Current GP methods may not optimally exploit reusable program components.

Purpose of the Study:

• To enhance building block identification in multidimensional genetic programming.
• To improve the efficiency and effectiveness of genetic programming through machine learning integration.

Main Methods:

• Investigated machine learning to bias the promotion of program components.
• Proposed two semantic operators, including a forward stagewise crossover, for strategic building block placement during crossover.
• Evaluated performance on regression problems and a large benchmark study.

Main Results:

• The forward stagewise crossover operator demonstrated significant improvements in regression problems.
• Achieved state-of-the-art results in a comprehensive benchmark study.
• Analyzed the propensity of architectures for heuristic search to utilize evolutionary information.

Conclusions:

• Machine learning integration and semantic crossover operators enhance genetic programming.
• The proposed methods offer a promising direction for improving evolutionary computation.
• Further analysis is needed on data representation collinearity and complexity for disentangling variation factors.