1

I'm looking to set up a linear regression using 2D Gaussian basis functions. My input training variables cover a two dimensional space. Before applying the machine learning (Bayesian linear regression), I need to select parameters for the Gaussians - mean and variance and also decide how many basis functions to use.

I am currently spacing the means (of a preallocated number of basis Gaussians) evenly over a grid, and just assuming constant variance. This is obviously not the best approach. Any ideas on how to calculate these variables?

benjanisa
  • 96
  • 1
  • 9
  • Not feeling competent enough to answer, but your issue is "model selection problem" - you may want to look it up. In general you shouldn't tune the parameters looking at the data - that spoils it, and inhibits generalization. – BartoszKP Nov 07 '13 at 11:29
  • see Tom Minka's answer: http://stackoverflow.com/questions/19824341/how-to-choose-gaussian-basis-functions-hyperparameters-for-linear-regression unfortunately there doesn't seem to be an established "best" way to do this – Yibo Yang Jan 14 '17 at 20:55

0 Answers0