Formation of Momentum and Learning Rate Profile for Online Training and Testing of HMLP with ALRPE

摘要

The technique of momentum and learning rate formation for children pedagogy was adopted to improve the performance of learning algorithm in training the hybrid multilayered perceptron network (HMLP) using modified recursive prediction error (MRPE) algorithm. An Adaptive Learning Recursive Prediction Error Algorithm (ALRPE) is proposed as a second version of MRPE with a guidance of a new profile of learning and momentum rate. An online model was used to forecast speed, revolution and fuel balanced in a Proton Gen2 car tank. The car measured the injected fuel from fuel injection sensor and became an input to the HMLP model to forecast the speed, revolution and fuel balanced in tank. These forecasted variables were also measured from the car sensors. To date, there is a restricted study on the effect of the profile of learning and momentum rate to the performance of HMLP. This study proposes a new profile of momentum and learning rate to improve the performance of the nonlinear modelling using HMLP. Previous conventional profile was developed only based on its general algorithm. The effect of the profile of momentum and learning rate to the generalisation of the HMLP using MRPE network was lack of discussion and thus, motivates this research using the proposed technique. Experimental results showed that the proposed ALRPE profile of momentum and learning rate can improved the performance of nonlinear HMLP model in the range of 0.002 dB to 3.15 dB of mean square error (MSE) in the model validation.