Presently, HamiltonianChain samples the total number of leapfrog steps used to produce the proposal from a pre-defined uniform distribution.
The No-U-Turn Sampler (NUTS) is a more efficient implementation of Hamiltonian Monte-Carlo which adaptively sets the number of steps taken per proposal.
NUTS should initially be implemented as a new class NutsChain (inheriting from HamiltonianChain)
NutsChain then needs to be bench-marked against HamiltonianChain in some test problems.
If it appears that there is no advantage to maintaining both NutsChain and HamiltonianChain, then the HamiltonianChain code will be replaced with that of NutsChain, which will then be removed.
Presently,
HamiltonianChain
samples the total number of leapfrog steps used to produce the proposal from a pre-defined uniform distribution.The No-U-Turn Sampler (NUTS) is a more efficient implementation of Hamiltonian Monte-Carlo which adaptively sets the number of steps taken per proposal.
NUTS should initially be implemented as a new class
NutsChain
(inheriting fromHamiltonianChain
)NutsChain
then needs to be bench-marked againstHamiltonianChain
in some test problems.If it appears that there is no advantage to maintaining both
NutsChain
andHamiltonianChain
, then theHamiltonianChain
code will be replaced with that ofNutsChain
, which will then be removed.