# \[AdvancedHMC\] Implementation of Custom HMCKernel
**URL:** https://discourse.julialang.org/t/advancedhmc-implementation-of-custom-hmckernel/132664
**Category:** Statistics
**Tags:** package
**Created:** [September 25, 2025, 5:51pm UTC](https://discourse.julialang.org/t/advancedhmc-implementation-of-custom-hmckernel/132664 "2025-09-25T17:51:21Z")
**Posts on this page:** 2
**Page:** 1
### Author:  [@sample\_slice\_of\_life](https://discourse.julialang.org/u/sample_slice_of_life)
#### Post date: [September 25, 2025, 5:51pm UTC](https://discourse.julialang.org/t/advancedhmc-implementation-of-custom-hmckernel/132664/1 "2025-09-25T17:51:21Z")
Hello all,
I am a student researching possible extensions of HMC-type samplers for settings in which posteriors exhibit high correlation between parameters.
I have a working implementation of a code for a toy problem in Julia using the AdvancedHMC.jl package. The results (posterior traceplots, convergence diagnostics, etc.) are _textbook_ – could not ask for better.
My issue: I have been attempting to replicate the same behavior in R. I am implementing both forward and adjoint sensitivity approaches to help obtain gradients w.r.t. parameters of the system, along with trying to implement NUTS, dual-averaging, and mass matrix adaptation (beyond diagonal). **Nothing I do gives results as efficient as Julia’s AdvancedHMC.jl or Stan in R.** Julia having access to AD methods obviously means the wall-clock time is better.
This issue is not ‘surprising’ – per se – but really dampens the exploration of any ideas in research because a suboptimal NUTS sampler means that any convergence diagnostics from testing an idea shroud any potential efficacy.
My Question: Is there any way to implement your _own_ HMCKernel instead of
> ϵ = find\_good\_stepsize(hamiltonian, collect(Float64, initial\_θ))
> integrator = JitteredLeapfrog(ϵ, 0.1)
> kernel = HMCKernel(Trajectory{MultinomialTS}(integrator, GeneralisedNoUTurn()))
> adaptor = StanHMCAdaptor(MassMatrixAdaptor(metric), StepSizeAdaptor(0.8, integrator))
and still be able to take advantage of, say, the StanHMCAdaptor and the like?
---
### Author:  [@Red-Portal](https://discourse.julialang.org/u/Red-Portal)
#### Post date: [October 15, 2025, 7:37am UTC](https://discourse.julialang.org/t/advancedhmc-implementation-of-custom-hmckernel/132664/2 "2025-10-15T07:37:31Z")
Short answer is yes, AdvancedHMC is highly modular and fairly easy to customize. But of course, how you would do that depends on what you want to customize specifically.