How to Improve Neural Network Training Using Evolutionary Algorithms

Pedro Carvalho, Nuno Lourenço, Penousal Machado

Abstract: Artificial neural networks are a staple of modern artificial intelligence. These systems must often undergo a training procedure to learn how to solve a designated task. Properly choosing and tuning an optimizer for a problem can significantly improve training speed and quality. Research into optimizers focuses on creating solutions that are generally applicable to any task, relying on parameter tuning for specialization. While parameter tuning is successful in adapting optimizers for a specific problem, the benefits of creating specialized optimizers from scratch is still underdeveloped. We propose an evolutionary framework called AutoLR, capable of evolving optimizers for specific tasks. We use the framework to evolve optimizers for a popular image classification problem and found that evolved optimizers are competitive with human-made optimizers. Furthermore, we find that the evolved solutions remain competitive when moved to a different dataset. Analysis of the best performing optimizers reveals that the system evolved novel behavior undiscovered by humans. The results achieved in this work suggest evolutionary algorithms can improve the quality of neural network training, motivating further research into the framework and its applications.