Yichen, Zhu, Chen Yang, and Adam Ghandar. “Evolving Neural Networks for Prediction with Negative Correlation Search: Application in Consumer Demand Forecasting.” In 2020 IEEE Symposium Series on Computational Intelligence (SSCI) , pp. 384-391. IEEE, 2020.
URL1
Neuroevolution is a powerful approach for learning neural networks for a large variety of machine learning applications. This paper describes a new approach that uses Negatively Correlated Search (NCS) to extend the basic NEAT neur. This approach uses NCS as a means for learning a population of negatively correlated neural network solutions, meaning that the individuals in the population are suited for performing well in different kinds of problem cases (or examples). An empirical evaluation of the proposed methodology we term NCS-NEAT leads to improved performance over basic NEAT in terms of accuracy and problem specific criteria and is a promising way to scale neuroevolution up to handle larger datasets without placing restrictions on the topology neural networks that are able to be learned. Two different machine learning problem classes were selected to be representative of a broad range of applications of deep learning and neuroevolution: the first is a price forecasting problem to predict the prices of mobile phones based on their characteristics; the second is a reinforcement learning test problem to validate the novel approach in a different type of problem.