Evolution strategies for weight optimization of Artificial Neural Network in time series prediction

• Published in: 2013 International Conference on Robotics, Biomimetics, Intelligent Computational Systems pp. 143 - 147
• Main Authors: Sulistiyo, Mahmud Dwi, Dayawati, Retno Novi, Nurlasmaya
• Format: Conference Proceeding
• Language: English
• Published: IEEE 01.11.2013
• Subjects: artificial neural network; Artificial neural networks; bacpropagation; Biological cells; evolution strategies; Neurons; order module; prediction; Prediction algorithms; Predictive models; Sociology; Statistics; time series
• DOI: 10.1109/ROBIONETICS.2013.6743594

Time series prediction is one of the problems to forecast an unseen value in the time series data. Basically, this technique uses a number of past sequential data values and then makes a calculation to yields a value for the next unseen data. The system using this technique requires a prediction model which had to be trained first. Artificial Neural Networks (ANN) is widely used as a prediction model for various time series prediction system. The training process of the ANN model is commonly done by using Back Propagation (BP) algorithm. But, there are many other alternative algorithms that can be applied to train the ANN model. One of them is by using an optimization algorithm. Evolution Strategies (ES) is one of the optimization algorithms that can be used as the training algorithm for the ANN prediction model. ES is one of the Evolutionary Algorithms (EAs) that has successfully solved many optimization problems. ES is expected to be faster in providing the optimal solution and in determining the optimal weights set for the prediction model. In this research, a prediction system is applied to forecast the order module of SLMA module: COS P / N S3081O-Q815-Xl15. The system is needed in order to predict the number of ordered modules, so that it can be used as consideration in the provision of modules. In the case of this research, it shows that the ES algorithm is able to be the alternative algorithm to train the ANN model in order to gain the optimal weights. It was shown in the training results that BP algorithm yields a bit better than ES algorithm. Meanwhile, in the testing results, the error yielded by the ES algorithm is less than the BP algorithm.