Thanks for the update, I’ve been using your package for my work since 1.3.0, and unfortunately I am still on 1.3 as the ABCDE method seems to do much better on my problem. I’ve been trying to come up with a MWE example for you but am not able to replicate the issue outside of my code.
I have looked into some other ABC packages, but it doesn’t seem there are any other julia packages that implement algorithms to handle ABC parameter estimation with ~30 variables, and I’d rather not move to python.
I can show some results from fitting here:
with KissABCv1.3
fakedist(x,y)=x+y #workaround for combined cost/simulate function you mentioned
plan = ABCplan(prior_dist, simulate, 0.0, fakedist)
res,_ = ABCDE(plan,16.0,nparticles = nparticles,generations = 0.5e3,earlystop = true,verbose=true,parallel = true)
with KissABCv3.0
res = smc(prior_dist,simulate; nparticles = 500,verbose = true,parallel = true,M = 10).P
using
approx_density = ApproxKernelizedPosterior(prior_dist,simulate,0.01)
res = sample(approx_density,AIS(30),MCMCThreads(),300,12;progress = true,ntransitions = 1600)
I have tried many parameter combinations to get the resulting distributions as tight as v1.3 with ABCDE, do you have any suggestions?
edit2: added examples