Neuroevolution: A Primer On Evolving Artificial Neural Networks

In the last years, artificial neural networks (ANN) have successfully been applied across a number of tasks, such as image classification, speech recognition and natural language understanding. Two main drivers, firstly a large amount of (labelled) data, and secondly growing compute resources, allowed to use technology that dates back to the 1960s and break records on many benchmarks. However, designing well-performing ANNs requires expert knowledge and experience. Neuroevolution aims at solving this difficult and often time-consuming process by applying evolutionary techniques. This article aims to provide an introduction to the re-emerging field of neuroevolution by supplying a definition, differentiating between other fields and areas and giving an overview by summarising a number of recent research papers. Inspired by biological neural networks, ANNs consist of many artificial neurons that are connected to each other. In a process, mostly referred to as training or learning, ANNs are optimised to solve specific, predefined tasks such as detecting faces. Learning is usually realised by employing a form of stochastic gradient descent (SGD) in order to change the network parameters in such a way that a future prediction for the same input is closer to the desired output. The architecture of a network, i.e. how the neurons are connected to each other, plays a very important role in whether or not an ANN can be trained to successfully learn a task. Over the years, human experts carefully designed complex architectures such as the VGGNet, AlexNet, GoogleNet, ResNet and many more to achieve and often surpass human-level performance on many different tasks.