I’m using the pm.gp.Latent.prior for a 2D function on polar-like coordinates (rho, theta). The X support (shape (n_points, 2)) of coordinates features angles -pi < theta < pi as the second dimension. Therefore, I need a kernel in the theta dimension which finds the closest angle difference between…

I think I saw a solution using a periodic distribution somewhere, maybe @aseyboldt or @bwengals knows?

Yes that’s the one I have in mind. Thanks @aseyboldt !

That post deals with a transform between polar and Cartesian coordinates. However, in my problem, I need to use the polar coordinates directly, because the values are expected to be correlated mostly along a constant rho “surface”. So transforming them to cartesian coordinates won’t help me (in fact…

The transformation is meant for cases where you work with the polar coordinates. The question is, which coordinates the sampler is supposed to see. If the sampler works with cartesian coordinates, then you can solve the problem of the wrapping around of angles, but you pay for this by introducing an…

I found some description of kernels for polar coordinates in this article Polar Gaussian Processes and Experimental Designs in Circular Domains . I’m thinking of implementing this kernel. My idea is that I’ll subclass Stationary, use its euclidean_dist and plug that into the formulas in the article…

Hmm, turns out I cannot simply use Stationary directly, because it uses the lengthscale in square_dist already. I need to do something like class PolarAngleChordal(pm.gp.cov.Stationary): """Kernel for polar angle coordinate using Wendland function of chordal distance Based on: Esperan Pad…

The chordal distance didn’t perform very well in comparison to the previous 3 compound kernels. The Geodesic distance-based kernel works much better. class PolarAngleGeodesic(pm.gp.cov.Stationary): """Kernel for polar angle coordinate using Wendland function of geodesic distance Based on:…

GP kernel for polar angle dimension

Questions

aseyboldt November 3, 2017, 1:29pm 3

@junpenglao I don’t know if that is what you mean, but we talked about angles and polar coordinates here.

Topic		Replies	Views
Haversine Kernel? v5 gaussian_process , modeling	6	245	June 26, 2024
GP Periodic Kernel - is pymc3 different from other sources? Questions	6	1091	June 30, 2020
Prediction with Latent GP using 2D kernel definition v5	2	225	June 28, 2024
Gaussian process lengthscale parameterization / trigger of compound sampler v5 gaussian_process , pytensor	3	894	June 1, 2023
Spherical coordinates Questions gaussian_process , modeling	0	297	August 30, 2022

GP kernel for polar angle dimension

Related topics