Open chonlei opened 3 years ago
On the re-run it doesn't look that great for the standard gaussian either. Worth investigating, I think, this is quite similar to what the bug in pop mcmc looked like
@ben18785 I've removed the banana test case, and I've run the correlated Gaussian for more iterations (doubled, 40k iterations).
@MichaelClerx What bug was it?
Thanks @chonlei -- I think @MichaelClerx means that the ECDFs don't look quite like the theoretical ones. I'm wondering a) whether the step size hyperparameter for this method is too short and b) whether just to run those two ECDF examples for much longer?
Thanks @chonlei -- I think @MichaelClerx means that the ECDFs don't look quite like the theoretical ones. I'm wondering a) whether the step size hyperparameter for this method is too short and b) whether just to run those two ECDF examples for much longer?
That's right! In pop mcmc we had a bug that caused something similar Plots showing the bug: #1238 PR fixing it #1246 The actual bug: https://github.com/pints-team/pints/pull/1246#issuecomment-760367408
@ben18785 @chonlei what's the status of this one?
This PR adds testing notebook for the Metropolis random walk. Closes #16.
So far metropolis works well for 2D Gaussian problem, but not for correlated Gaussian and banana problems. I guess it is somewhat expected given this is a vanilla Metropolis MCMC method. I suppose correlated problems are not easy for such a vanilla method, never mind the banana problem.
I am happy for this method to behave in such a way, though let me know if any one thinks we should try to make it works for the other problems as well.