Auto-Adaptive the Weight In Batch Back Propagation Algorithm Via Dynamic Learning Rate

Abstract

Batch back propagation (BBP) algorithm is commonly used in many applications, including robotics, automation, and global positioning systems. The man drawbacks of batch back propagation (BBP) algorithm is slow training, and there are several parameters needs to be adjusted manually, also suffers from saturation training. The objective of this study is to improve the speed uptraining of the BBP algorithm and to remove the saturation training. To overcome these problems, we have created a new dynamic learning rate to escape the local minimum, which enables a faster training time. We presented dynamic batch backpropagation algorithm (DBBPLR) which training with dynamic learning rate. This technique was implemented using a sigmoid function. The XOR problem, the Balance dataset, and the Iris dataset were used as benchmarks with different structures to test the efficiency of the dynamic learning rate. The real datasets were divided into a training set and a testing set, and 75 experiments were carried out using Matlab software2016a. From the experimental results, it can be shown that the DBBPLR algorithm provides superior performance over the existing BBP algorithm in terms of training, the speed of training, time training, number of epochs and accuracy training and also with existing work.