Section 3.2: I know we might have some (friendly!) differences over here with Dan in particular, but it might be good to mention that optimizing the GPs together with certain physical models out there might not be easy. In particular, when using samplers, it’s very important to keep in mind that for MCMC based inference the starting points can have a huge impact as to whether your fit converges or not. This is particularly true for some hard problems in radial-velocity data, where the GP hyperparameters might fall close to planetary parameters (e.g., GP period and planetary periods), or where multi-modality might be present. I’ve found a lot of success with Nested Samplers on this because of their thorough (although costly!) exploration of the parameter space, as they sample from the prior. All this is to say: might be good to mention Nested Samplers too as a good alternative for sampling here (I know they are mentioned in Section 3.3 --- but they are only mentioned as tools to get the log-evidence)
dfm
commented
2 years ago
Section 3.2: I know we might have some (friendly!) differences over here with Dan in particular, but it might be good to mention that optimizing the GPs together with certain physical models out there might not be easy. In particular, when using samplers, it’s very important to keep in mind that for MCMC based inference the starting points can have a huge impact as to whether your fit converges or not. This is particularly true for some hard problems in radial-velocity data, where the GP hyperparameters might fall close to planetary parameters (e.g., GP period and planetary periods), or where multi-modality might be present. I’ve found a lot of success with Nested Samplers on this because of their thorough (although costly!) exploration of the parameter space, as they sample from the prior. All this is to say: might be good to mention Nested Samplers too as a good alternative for sampling here (I know they are mentioned in Section 3.3 --- but they are only mentioned as tools to get the log-evidence)