mauricelanghinrichs / abcdez.jl Goto Github PK

implements #5, to be released with v0.2.0 as new main feature; does not break previous code, but induces changed behaviour. Kmcmc steps may early exit due to new Kmcmc_min; output of abcdesmc! method now additionally contains Kmcmcs info.

@JuliaRegistrator register()

Look into the possibility to set a high Kmcmc keyword argument in abcdesmc! method, with an option (new keyword argument) to early exit the MCMC iterations when a sufficiently high number of acceptances (as provided/desired) has already been reached.

This may help to diversify particles better at the last epsilon steps (higher Kmcmc), without adding too much unnecessary computation load at the first iterations (using early exit). See into smc at KissABC where something similar exists.

Potentially also write the actual number of used MCMC steps into the results object (if verboseout=true).

• Addition of smooth and strict kernels (Epa, Indicator), including new tests.
• Small change to make eps choice non-increasing in abcdesmc! method (no changed behaviour with previous default indicator kernel).

@JuliaRegistrator register()

This issue is used to trigger TagBot; feel free to unsubscribe.

If you haven't already, you should update your TagBot.yml to include issue comment triggers.
Please see this post on Discourse for instructions and more details.

If you'd like for me to do this for you, comment TagBot fix on this issue.
I'll open a PR within a few hours, please be patient!

Release to update compat helper suggested StatsBase.jl dependency change.

@JuliaRegistrator register()

below a collection of potential improvements for future releases.

• maybe turn off the default tuning of the proposal in the SMC method (shrinkage of gamma parameter). for large priors I noticed that the acceptance ratio can be low initially, causing a proposal shrinkage way too early. otherwise the shrinkage works nicely, but the default should be conservative.
• maybe use the strict indicator kernel as default (IndicatorStrict0toϵ) for the SMC method. without strictness (i.e. with the current default Indicator0toϵ kernel) the SMC alg. sometimes gets stuck if particle numbers are on the lower side. disadvantage is that the non-strict kernel is better for small discrete problems where perfect convergence may be reached with ABC; but this case is rare in practice, hence I could just document the kernel choice more for directing users appropriately and switch to the strict kernel.
• the strict kernel in the SMC method (see above, potential new default) will at some point kill all particles if eps is too low for the natural noise level (at posterior) of the model simulations. this is all fine in theory, but the method should handle these cases with a warning (or so) and still return the empty result without errors. currently it errors which can be annoying and might confuse users.
• check again the base "particle operations" that were taken over from KissABC.jl for potential type instabilities of the tuple handlings. in case there are, check if they are relevant for performance. think about new implementation; should not affect user API much / at all.

any further suggestions are very very welcome!

@JuliaRegistrator register()