More consistent learnt relationships in regression neural networks using Fit to Median Error measure

Parkes, Amy I., Sobey, Adam J. and Hudson, Dominic A. (2025) More consistent learnt relationships in regression neural networks using Fit to Median Error measure. Engineering Applications of Artificial Intelligence, 144, [110113]. (doi:10.1016/j.engappai.2025.110113).

Record type: Article

Abstract

Machine learning is increasingly used to optimise complex systems. The quality of this optimisation is dependent on the models having a reasonable physical representation of the system. However, standard error metrics are pointwise, providing a reasonable physical representation only under certain constraints. As an example, data models of ship powering have low errors values, <2%, but fail to consistently approximate the input–output relationships and cannot be used to optimise performance. This paper illustrates that the Fit to Median error measure can be used to assess how well the ground truth is modelled. It provides more consistent learnt relationships and improves the extrapolation accuracy of neural networks. This is illustrated on real-world data used for ship power prediction.

Text

revised manuscript 2 - Accepted Manuscript

Restricted to Repository staff only until 28 January 2027.

Available under License Creative Commons Attribution Non-commercial No Derivatives.

Request a copy