search

jamesrobertlloyd / gp-structure-search

MIT License
229 stars 84 forks source link

Nyquist aliasing with periodic kernels - and defaults #16

Closed jamesrobertlloyd closed 11 years ago

jamesrobertlloyd commented 11 years ago

Prevent periods smaller than smallest data distance to prevent trivial aliasing

In general, what are 'sensible' default values

jamesrobertlloyd commented 11 years ago

This is fixed (mildly hackily) - but it has raised another issue. What are sensible default values for random restarts of the period of periodic kernels? Problem is that there are lots of local optima at harmonics. I've have inserted a hack of centering the period on -2 for initialisation (the assumption of ~ 10 wiggles?) but this needs some more serious thought. Really, we need a prior on the structure of data sets!

jamesrobertlloyd commented 11 years ago

Currently have hacked in that no period starting value can be smaller than 5 * the smallest inter data gap - this is a magic number but is hopefully vaguely reasonable.

duvenaud commented 11 years ago

Seems to work!

jamesrobertlloyd commented 11 years ago

Improved the heuristic so that it also disallows periods smaller than m * the average distance between data points. This makes the heuristic still work when the data is irregularly spaced.