Identification of chaotic systems with large noise based on regularized feedforward neural networks

A method for identification of chaotic systems with large noise based on regularized feedforward neural networks is proposed. The regularization method can improve greatly the generalization performance of the feedforward networks. At various noise levels, we train feedforward networks with regularization parameter and clarify fundamental properties of regularized feedforward networks to learn noisy chaotic systems by some numerical experiments. We also evaluate the identified models with reconstruction of attractors by the identified models. Simulations show that the identified models can approach to original chaotic systems and extract dynamical characteristics of original chaotic systems.