Models of performance of evolutionary program induction algorithms based on indicators of problem difficulty.

Area of Science:

• Computer Science
• Artificial Intelligence
• Algorithm Theory

Background:

• Evolutionary algorithm theory aims to model algorithm behavior for practical problem-solving.
• Previous performance estimation methods for evolutionary program-induction algorithms (EPAs) had limitations including reliance on reference problems, manual feature selection, and opaque models.

Purpose of the Study:

• To develop an improved technique for estimating EPA performance that overcomes the limitations of prior methods.
• To create more general, simpler, and accurate performance models for EPAs.

Main Methods:

• Introduced a novel set of general features based on finite differences to assess problem difficulty for EPAs.
• Applied the new technique to symbolic regression, Boolean function induction, and wind speed forecasting problems.
• Modeled various EPAs and autoregressive models.

Main Results:

• The new method produced significantly simpler and more accurate performance models compared to previous approaches.
• Models for wind speed forecasting using both EPAs and autoregressive models outperformed previous models.
• Demonstrated the technique's generality beyond EPAs.

Conclusions:

• The proposed finite difference-based feature set offers a practical and effective approach to modeling EPA performance.
• This improved technique provides more accessible and reliable guidelines for algorithm selection in real-world applications.
• The method's applicability extends to other domains like time-series forecasting.