Automatic tuning of DR-G-HMC parameters

Develop an automatic tuning procedure for the parameters of the Delayed Rejection Generalized Hamiltonian Monte Carlo (DR-G-HMC) algorithm, with particular attention to the damping parameter γ that governs momentum conservation during partial momentum refreshment.

Background

DR-G-HMC combines delayed rejection with generalized Hamiltonian Monte Carlo to enable dynamic, per-iteration step-size selection and improve sampling efficiency on multiscale target distributions. Its performance depends on several tunable hyperparameters, notably the damping factor γ that controls the degree of momentum refreshment and thus directed motion.

While NUTS provides effective auto-tuning for HMC and the MEADS algorithm offers tuning for a non-reversible GHMC variant, no analogous automatic tuning procedure has been established for DR-G-HMC. The authors indicate that solving this would likely yield substantial performance improvements, similar to how NUTS improves upon HMC.