figure out a robust way to optimize hyper-parameters

[reedessick]

This will be needed to get the error regions on the process accurately. We also need to make sure this is quite robust. Preferably, this should happen automatically, without human input or fine-tuning.

[reedessick]

write visualization for marginalization over hyper-parameters as part of investigate_hyperparameters

• attempt to draw from a direct sum over many weighted Gaussian Processes.
• compute the mean and covariance of the to "total" process by weighted sums over the individual processes' means and covariances

See if these two approaches agree.

check how polyfit affects the optimal choice of hyperparameters. Should it? If it does, do the resulting residuals look reasonable? That gets back to the visualization engine that's needed

[reedessick]

it looks like the mcmc sampling of the hyperparameter "posteriors" works well, but we need to include things like stripping the burn in automatically and possibly downsampling based on autocorrelation length estimates.