Deep Neuroevolution: Training Neural Networks Using a Matrix-Free Evolution Strategy

In this paper, we discuss an evolutionary method for training deep neural networks. The proposed solution is based on the Differential Evolution Strategy (DES) – an algorithm that is a crossover between Differential Evolution (DE) and the Covariance Matrix Adaptation Evolution Strategy (CMA-ES). We combine this approach with Xavier’s coefficient-based population initialization, batch processing, and gradient-based mutations—the resulting weight optimizer is called neural Differential Evolution Strategy (nDES). Our algorithm yields results comparable to Adaptive Moment Estimation ADAM for a convolutional network training task (50K parameters) on the FashionMNIST dataset. We show that combining both methods results in better models than those obtained after training by either of these algorithms alone. Furthermore, nDES significantly outperforms ADAM on three classic toy recurrent neural network problems. The proposed solution is scalable in an embarrassingly parallel way. For reproducibility purposes, we provide a reference implementation written in Python.