Possible Abstract mismatch for the 'tuning' argument

[kangkang-Zhao]

line 79 in ext/ahmc_impl/ahmc_config_impl.jl, the argument of 'tuning' is supposed to have the abstract type of "HMCTuning", passed from 'proposal_tuning', but here the type "StanLikeTuning" has the parent abstract as "MCMCTransformTuninng", perhaps there is a conflict.

[oschulz]

@Micki-D could you take a look?

[Micki-D]

I'll check it out!

[Micki-D]

Hi @kangkang-Zhao, this method for ahmc_adaptor() that you reference is supposed to be used for the transformation tuner, so there is no conflict.
That being said, this method is technically redundant because the BAT tuners separate the proposal tuning and transformation tuning that is rolled into one in the AdvancedHMC package.

Because we separate the two tunings, we only pass a StepSizeAdaptor to AdvancedHMC and do the mass matrix tuning that would be performed by AdvancedHMC inside of BAT instead, via the tuning of the adaptive_transform. This is consistent with the AdvancedHMC tuning, as we update the HMC proposal's hamiltonian with the function set_proposal_transform!!() here.

We decided to keep these methods for ahmc_adaptor() to offer more functionality for experimentation and to have the option of replicating the exact behavior of AdvancedHMC's tuners.