Optimizers parameter space probing and not finding solution?

[lancelot49]

Hello Kyle,

I am trying to fit and optimize a RV data, I know the period and the eccentricity of the system, however the rv geometry, pushes the eccentricity much higher, is there any way to run the estimator with a fixed eccentricity? Not my main problem however, I manually guessed my way to a okay fit (picture below), but then trying to optimize with nelder for asini, q, vgamma and t0, never results in a solution, let it run for close to three hours yesterday but it never converged. How does nelder mead probe the parameter space? And is there any reason why it does not find a solution within few thousand iterations? Also, from what i see, when optimizing the rv, the tutorials online never include p0, even tough, changes to p0 slightly affect the rv?

Any help will be greatly appreciated,
cheers!

[aprsa]

Not Kyle, but let me chime in anyway. :) What do you mean exactly by "never converges?" Do the parameter values not change, or do they get stuck at a local solution, do they push against the parameter limits, or something else entirely? If, for example, your asini wants to converge to negative values, that is the sign that primary and secondary RVs might be switched.

[lancelot49]

Had to zip it since github does not support the .bundle type, this was saved, as you said, just before I ran run_solver.
before_run_solver.zip

[aprsa]

I managed to reproduce this and I didn't see anything glaringly obvious that would cause -infs. Please stay tuned, we're looking into this.

[kecnry]

Thanks for attaching the bundle, that helps a lot in tracking down what is happening. Your solver is using the nm_fit compute options which fails when running b.run_compute(compute='nm_fit') with the following error:

ValueError: Could not compute intensities with atm='ck2004'.  Try changing atm to a table that covers a sufficient range of values (or to 'blackbody' in which case ld_mode will need to be set to 'manual' and coefficients provided via ld_coeffs).  Enable 'warning' logger to see out-of-bounds arrays.

Your phoebe01 compute options use atm='blackbody', which do succeed, which is probably why this was easily missed. If you set the atmospheres to blackbody in compute='nm_fit', then they do have finite likelihoods and the optimizer does run.

With the light curve disabled and the RVs dynamical, this probably should not need to be called and can be fixed to avoid the error (I created #812 to track that problem). For now, so long as you're only fitting the RVs, you can get around that by setting the atmospheres to blackbody. But once you want to fit the light curves as well, I'd suggest re-enabling the atmospheres and ensuring that you're in a physical region of parameter space. Currently your stars both have radii of 1 solRad, but the primary has a mass of 28 solMass and a temperature of 38000 K, which falls outside the bounds of the atmosphere table.

[lancelot49]

Thank you so much, just tried a few iteration run at it worked, would have never figured this out with your guys help!

[kecnry]

Just to follow-up, the bug with dynamical RVs failing with an atmosphere-out-of-bounds error has been fixed in #823 and will be included in the next (2.4.13) bugfix release.