Time Consuming Numerical Model Calibration Using Genetic Algorithm (GA), 1-Nearest Neighbor (1NN) Classifier and...

Area of Science:

• Computational Science
• Optimization Algorithms
• Machine Learning Applications

Background:

• Single Objective Genetic Algorithm (SGA) optimization requires extensive objective function evaluations, often hindering real-world numerical model calibration.
• Evaluating solution quality in these calibration problems is computationally intensive and time-consuming.

Purpose of the Study:

• To present an effective and efficient dynamic approximation model, SGA-1NN-PCA, to decrease the number of required fitness evaluations in SGA.
• To enhance the speed and efficiency of numerical model calibration using genetic algorithms.

Main Methods:

• The SGA-1NN-PCA algorithm utilizes a 1-Nearest Neighbor (1NN) classifier to predict objective function values, reducing the need for direct evaluations.
• Principal Component Analysis (PCA) is employed to reduce the dimensionality of optimization parameters, further saving computational time for the 1NN classifier.
• Training data for the 1NN classifier is generated from the early stages of the optimization process.

Main Results:

• The proposed SGA-1NN-PCA method demonstrated a significant reduction in computational load.
• The algorithm required only approximately 25% of the actual fitness evaluations compared to the standard SGA.
• This efficiency gain is crucial for complex and time-consuming numerical model calibration tasks.

Conclusions:

• SGA-1NN-PCA offers a computationally efficient alternative for optimizing numerical models using genetic algorithms.
• The integration of 1NN and PCA effectively approximates objective function values, drastically reducing evaluation requirements.
• This approach is particularly beneficial for real-world problems where fitness evaluations are a bottleneck.