Abstract

Artificial Neural Network (ANN) algorithms are often used to predict the output according to the input given to it. Before being implemented, ANN requires a number of exercises (training) to obtain the appropriate weight, so that the difference between the output and the set target is as minimal as possible. The training process requires a lot of training data and the number of repeated trainings reaches hundreds, thousands or even millions of times. Therefore, this training requires a large memory and a very long time. Given these conditions, ANN cannot be directly implemented in a system that requires real-time data processing, with the speed of time and accuracy of the output. Especially when applied to vehicle systems that definitely demand speed and execution accuracy when encountering changes in sensor parameters embedded in it, so that unwanted conditions can be avoided. For this reason, training can be done first to obtain network weights, then these weights are injected into the system which requires real-time execution.