A Novel Evolutionary Approach for Neural Architecture Search

Convolutional Neural Networks (CNNs) have proven to be an effective tool in many real-world applications. The main problem of CNNs is the lack of a well-defined and largely shared set of criteria for the choice of architecture for a given problem. This lack represents a drawback for this approach since the choice of architecture plays a crucial role in CNNs’performance. Usually, these architectures are manually designed by experts. However, such a design process is computationally intensive because of the trial-and-error process and also not easy to realize due to the high level of expertise required. Recently, to try to overcome those drawbacks, many techniques that automize the task of designing the architecture neural networks have been proposed. To denote these techniques has been defined the term “Neural Architecture Search” (NAS). Among the many methods available for NAS, Evolutionary Computation (EC) methods have recently gained much attention and success. In this paper, we present a novel approach based on evolutionary computation to optimize CNNs. The proposed approach is based on a newly devised structure which encodes both hyperparameters and the architecture of a CNN. The experimental results show that the proposed approach allows us to achieve better performance than that achieved by state-of-the-art CNNs on a real-world problem. Furthermore, the proposed approach can generate smaller networks than the state-of-the-art CNNs used for the comparison.