Sliding-window learning using MLP networks with data store management

This paper explore the performance of sliding-window based for training multilayer perceptron neural network with correlated data. Online learning is usually employed when system variables are time varying. It is also used when it is not suitable to obtain a full history of offline data about the system as compared to offline learning. Sliding-window framework is proposed to combine the robustness of offline learning with the ability of online learning to track time varying elements of the process under investigation. This paper evaluates the performance of first order back propagation, second order conjugate gradient algorithms and the recent binary ensemble training algorithms with sliding-window learning routine. Different data store management techniques are presented to deal with the correlation problem.