Constrained Drawing Functions
For the problem of drawing uniformly from the prior, given a minimum likelihood,
these are the currently implemented methods:
Check the source of this Constrainer for reference, and as a starting point to
implementing your own method.
-
class nested_sampling.samplers.rejection.RejectionConstrainer[source]
Bases: object
Simplest draw function using rejection sampling.
Guaranteed to yield uniform points and not to exclude any maxima.
Least efficient method.
-
class nested_sampling.samplers.svm.SVMConstrainer[source]
Bases: object
This constrainer uses probabilistic Support Vector Machines classifier
with Radial basis functions (sklearn.svm.NuSVC)
to find a border separating live points and already discarded points.
Then, points are filtered using this classifier. Only points matching
the classifier are evaluated.
-
class nested_sampling.samplers.friends.FriendsConstrainer(rebuild_every=50, radial=True, metric='euclidean', jackknife=False, force_shrink=False, hinter=None, verbose=False)[source]
Bases: object
Rejection sampling pre-filtering method based on neighborhood to live points.
“Distant” means in this implementation that the distance to a cluster member
is large.
The maximum distance to a cluster is computed by considering each
cluster member and its k nearest neighbors in turn, and
computing the maximum distance.
| Parameters: |
- rebuild_every – After how many iterations should the clustering
distance be re-computed?
- radial – if radial = True, then the normal euclidean distance is used.
otherwise, the absolute coordinate difference in each dimension is used.
- metric – metric to use. Use ‘chebyshev’ for SupFriends, in which case then
the supremum norm is used. Use ‘euclidean’ for RadFriends, via
the euclidean norm.
- jackknife – if True, instead of leaving out a group of live points in
the distance estimate, only one is left out in turn (jackknife resampling
instead of bootstrap resampling).
- force_shrink – if True, the distance can only decrease between sampling steps.
|
|---|
-
are_inside_rect(u)[source]
Check if the new points are near or inside one of our clusters
-
cluster(u, ndim, keepRadius=False)[source]
-
is_inside(u)[source]
Check if this new point is near or inside one of our clusters
-
class nested_sampling.samplers.mcmc.BaseProposal(adapt=False, scale=1.0)[source]
Bases: object
Base class for proposal function.
| Parameters: |
- scale – Scale of proposal
- adapt – Adaptation rule to use for scale, when new_chain is called.
|
|---|
If adapt is False, no adaptation is done. If adapt is ‘Sivia’, the rule
of Sivia & Skilling (2006) is used. If adapt is something else,
a crude thresholding adaptation is used to gain ~50% acceptance.
-
class nested_sampling.samplers.mcmc.GaussProposal(adapt=False, scale=1.0)[source]
Bases: nested_sampling.samplers.mcmc.BaseProposal
Symmetric gaussian proposal.
@see BaseProposal
-
class nested_sampling.samplers.mcmc.MCMCConstrainer(nsteps=200, proposer=MultiScaleProposal(loc=-4.5, scale=1.5, adapt=False), nmaxsteps=10000)[source]
Bases: object
Markov chain Monte Carlo proposals using the Metropolis update:
Do a number of steps, while adhering to boundary.
-
class nested_sampling.samplers.mcmc.MultiScaleProposal(loc=-4.5, scale=1.5, adapt=False)[source]
Bases: nested_sampling.samplers.mcmc.BaseProposal
Proposal over multiple scales, inspired by DNest.
Uses the formula
\(x + n * 10^{l - s * u}\)
where l is the location, s is the scale and u is a uniform variate,
and n is a normal variate.
@see MultiScaleProposal
-
class nested_sampling.samplers.galilean.GalileanRadFriendsConstrainer(nlive_points, ndim, velocity_scale=0.5, nsteps=20, plot=False)[source]
Bases: nested_sampling.samplers.galilean.HybridRadFriendsConstrainer
Galilean Nested Sampling is a special case of Nested Sampling using Hamiltonian Monte Carlo.
Here we use RadFriends to determine the reflection surfaces.
-
are_inside_rect(u)
Check if the new points are near or inside one of our clusters
-
is_inside(u)
Check if this new point is near or inside one of our clusters
-
reflect(u, v)[source]
Find reflection surface for line u + t*v, t>0
-
class nested_sampling.samplers.galilean.HybridRadFriendsConstrainer(nsteps=20, plot=False)[source]
Bases: object
Base class to use RadFriends (with enforced shrinking, JackKnife resampling)
in combination with local step sampling methods.
-
are_inside_rect(u)[source]
Check if the new points are near or inside one of our clusters
-
is_inside(u)[source]
Check if this new point is near or inside one of our clusters
-
class nested_sampling.samplers.galilean.MCMCRadFriendsConstrainer(proposal_scale=3, nsteps=20, plot=False)[source]
Bases: nested_sampling.samplers.galilean.HybridRadFriendsConstrainer
Markov Chain Monte Carlo sampling
Here we use RadFriends to restrict the proposal distribution.
-
are_inside_rect(u)
Check if the new points are near or inside one of our clusters
-
is_inside(u)
Check if this new point is near or inside one of our clusters