Controlling Individuals Growth in Semantic Genetic Programming through Elitist Replacement

Area of Science:

• Computer Science
• Artificial Intelligence
• Machine Learning

Background:

• Geometric semantic genetic operators offer unimodal error surfaces for supervised learning.
• A key drawback is the generation of large, uninterpretable data models.

Purpose of the Study:

• To reduce the size of data models generated by geometric semantic operators.
• To maintain the advantage of unimodal error surfaces while improving model interpretability.

Main Methods:

• An elitist strategy was implemented where offspring are accepted only if they possess superior fitness to their parents.
• The proposed method was tested on five complex real-life supervised learning problems.

Main Results:

• The elitist geometric semantic operators achieved comparable fitness results to standard operators.
• The generated data models were significantly smaller and more manageable.

Conclusions:

• Elitist geometric semantic operators offer a viable solution to the model size issue in genetic programming.
• This approach enhances the interpretability of models generated through supervised learning.