Low-Overhead Learning: Quantized Shallow Neural Networks at the Service of Genetic Algorithm Optimization

Area of Science:

• Computational Intelligence
• Machine Learning
• Optimization Algorithms

Background:

• Online parameter tuning improves optimization algorithms like genetic algorithms (GAs) by adjusting mutation and crossover rates.
• Existing methods face high computational costs and poor adaptability in dynamic environments, especially with machine learning integration.

Purpose of the Study:

• To propose a quantized shallow neural network (SNN) for efficient, dynamic adjustment of GA mutation and crossover rates.
• To reduce computational overhead and enhance adaptability in complex fitness landscapes.

Main Methods:

• A quantized SNN was developed as a learning-based component for GA parameter tuning.
• Quantization techniques, including Quantization-aware Training (QaT) and Post-training Quantization (PtQ), were applied.
• Runtime-generated data was utilized for training and adaptation.

Main Results:

• The quantized SNN achieved high-quality solutions on 15 continuous benchmark functions.
• Significant reductions in execution time were observed compared to other shallow learning methods.
• The approach demonstrated a balance between computational efficiency and solution performance.

Conclusions:

• Quantized SNNs offer an efficient solution for dynamic parameter tuning in GAs.
• This method enhances the applicability of shallow learning in complex optimization tasks.
• The proposed approach effectively reduces computational burden while maintaining competitive performance.