A Study on the Scalability of Artificial Neural Networks Training Algorithms Using Multiple-Criteria Decision-Making Methods

In recent years, the unrestrainable growth of the volume of data has raised new challenges in machine learning regarding scalability. Scalability comprises not simply accuracy but several other measures regarding computational resources. In order to compare the scalability of algorithms it is necessary to establish a method allowing integrating all these measures into a single rank. These methods should be able to i) merge results of algorithms to be compared from different benchmark data sets, ii) quantitatively measure the difference between algorithms, and iii) weight some measures against others if necessary. In order to manage these issues, in this research we propose the use of TOPSIS as multiple-criteria decision-making method to rank algorithms. The use of this method will be illustrated to obtain a study on the scalability of five of the most well-known training algorithms for artificial neural networks (ANNs).