Order-Based Schedule of Dynamic Topology for Recurrent Neural Network

It is well-known that part of the neural networks capacity determined by their topology and employed training process. How a network should be designed how it updated every time new data acquired, an issue remains open since its usually limited to process trial error, based mainly on experience designer. To address this issue, algorithm provides plasticity recurrent (RNN) applied series forecasting proposed. A decision-making grow prune paradigm created, calculation data’s order, indicating in which situations during re-training (when received), increase or decrease connections, giving as result dynamic architecture can facilitate design implementation network, well improve behavior. The proposed was tested with some M4 competition, using Long-Short Term Memory (LSTM) models. Better results were obtained for most tests, models both larger smaller than static versions, showing average improvement up 18%.