Genetic-algorithm-optimized neural networks for gravitational wave classification

Gravitational-wave detection strategies are based on a signal analysis technique known as matched filtering. Despite the success of filtering, due to its computational cost, there has been recent interest in developing deep convolutional neural networks (CNNs) for detection. Designing these remains challenge most procedures adopt trial and error strategy set hyperparameter values. We propose new method optimization genetic algorithms (GAs). compare six different GA variants explore choices GA-optimized fitness score. show that can discover high-quality architectures when initial seed values far from good solution well refining already networks. For example, starting architecture proposed by George Huerta, network optimized over 20-dimensional space 78% fewer trainable parameters while obtaining an 11% increase accuracy our test problem. Using algorithm refine existing should be especially useful if problem context (e.g., statistical properties noise, model, etc) changes one needs rebuild network. In all experiments, we find discovers significantly less complicated compared network, suggesting it used prune wasteful structures. While have restricted attention CNN classifiers, applied within other machine learning settings.