Time series prediction with evolvable block-based neural networks

This paper presents a time series prediction technique using the block-based neural networks (BbNNs). Building a model dynamical system can be a general approach to the time series prediction problem. However, the functional form and the order of the dynamics of the process generating the time series data are usually unknown. BbNNs, an evolvable neural network model with simultaneous optimization of network structure and connection weights by use of evolutionary algorithms, provide a model-free estimation of underlying nonlinear dynamical systems. Empirical results with a benchmark Mackey-Glass time series show that the evolved BbNNs can predict the future behavior of a complex dynamical system with sufficient accuracy.