Abstract
The process of choosing appropriate machine learning parameter values for a domain is difficult as there are often many parameters involved in an algorithm. This is the motivation for automated machine learning, which involves the data-driven choice of parameter values. While there are state-of-the-art methods for this process, the simple method of grid search is often used despite being computationally expensive. Also, all AutoML methods including grid search are performed by trying to choose parameter value combinations that minimise aggregate total error.
However, aggregate error does not illustrate the behaviour of a learning algorithm for the examined parameter values. Grid search based on total error is a black box process with no understanding of why a certain set of parameters have been chosen. Instead, decomposing the prediction error provides a greater understanding of the behaviour of an algorithm and its parameters. In this paper, the error associated with grid search is decomposed for a number of machine learning algorithms. This shows that there are many parameter value combinations that exhibit similar total error but actually provide different behaviour by targeting a different error component.