CNN-HT: A Two-Stage Algorithm Selection Framework.

Area of Science:

• Computer Science
• Artificial Intelligence
• Optimization

Background:

• The No Free Lunch Theorem highlights the need for problem-specific algorithm selection.
• Existing single-stage algorithm selection methods often require complete retraining for new algorithm combinations.

Purpose of the Study:

• To introduce the CNN-HT, a novel two-stage algorithm selection framework.
• To improve algorithm selection by leveraging Convolutional Neural Networks (CNN) for problem classification and Hypothesis Testing (HT) for algorithm recommendation.

Main Methods:

• Utilized Exploratory Landscape Analysis (ELA) features as input for problem classification.
• Employed Convolutional Neural Networks (CNN) for initial problem classification, followed by Hypothesis Testing (HT) for algorithm selection.
• Implemented feature selection techniques to optimize the classification model.

Main Results:

• Achieved 96% average accuracy in problem classification using CNN, outperforming Random Forest and Support Vector Machines.
• Increased classification accuracy to 98.8% after feature selection, enhancing performance and reducing computational cost.
• Demonstrated superior performance of the CNN-HT framework with better average rankings compared to individual algorithms and other combination approaches.

Conclusions:

• The CNN-HT framework provides an effective and adaptable solution for algorithm selection in optimization.
• The two-stage approach allows for modifications without full model retraining, offering a significant improvement over single-stage methods.
• The high accuracy achieved in problem classification validates the efficacy of the first stage of the CNN-HT method.