Technique of learning rate estimation for efficient training of MLP

A new computational technique for training of multilayer feedforward neural networks with sigmoid activation function of the units is proposed. The proposed algorithm consists two phases. The first phase is an adaptive training step calculation, which implements the steepest descent method in the weight space. The second phase is estimation of calculated training step rate, which reaches a state of activity of the units for each training iteration. The simulation results are provided for the test example to demonstrate the efficiency of the proposed method, which solves the problem of training step choice in multilayer perceptrons.