User-specified resampling methods

[pkairys]

I'm creating this issue to discuss user-specified resamplers. This would be useful to specify different resamplers for different applications. For example when the posterior is multimodal. ]. It would also help to optimize the resampling methods separately from the OBE classes in order to increase performance.

Thoughts?

[rmcmichael-nist]

Hi @pkairys. I welcome the discussion.

With the code as is, a user can create a child of OptBayesExpt and overwrite the resample function, but maybe that's not what you're suggesting.

If the goal is to add alternative, user-selectable resamplers to a new version of the distribution, then I'd suggest adding alternative resampling methods to the ParticlePDF class and select between them with a keyword argument in init()

Would there be an advantage to putting the resample function in a separate Resampler class that would be inherited by ParticlePDF? I suppose that if we also put (e.g.) the resample_test function in a separate class, the two functions could be developed in parallel without conflct. Admittedly, I'm mostly ignorant of the branch & pull functions of git so if I'm missing something, please let me know.

[pkairys]

My idea was in line with your second point. Basically, allow a user to specify a resampler function when initializing the particlePDF. This can either be done by passing a keyword argument to specify which resampler. Or equally, by passing a resampler function to the class.

Something like

def custom_resampler(weights, particles, params):
    ....
    return new_particles

custom_params= {'a':0.98, ...}

mypdf = ParticlePDF(..., resampler=custom_resampler, resampler_params = custom_params)

I think this would be really useful because it would allow me to optimize my resampler implementation as needed for the application without having to modify any OBE classes or code.

[rmcmichael-nist]

OK, yes. Go ahead and make your version of ParticlePDF with extra custom_reampler methods and the resampler keyword. It's OK to coach/prompt/nag me on use of branching and pulling if you need those features.