A method for exploiting domain information in astrophysical parameter estimation

Abstract

I outline a method for estimating astrophysical parameters (APs) from multidimensional data. It is a supervised method based on matching observed data (e.g. a spectrum) to a grid of pre-labelled templates. However, unlike standard machine learning methods such as ANNs, SVMs or k-nn, this algorithm explicitly uses domain information to better weight each data dimension in the estimation. Specifically, it uses the sensitivity of each measured variable to each AP to perform a local, iterative interpolation of the grid. It avoids both the non-uniqueness problem of global regression as well as the grid resolution limitation of nearest neighbours.