SampleBelief(samples, weights)represents a distribution by weighted particles: the general-purpose fallback when no closed form is available, and the natural output of a sampler.
Sources: code:
open_model.jl,messages.jl; Doucet, de Freitas & Gordon (eds.), Sequential Monte Carlo Methods in Practice (Springer 2001)Bibliography: Doucet et al. 2001
Theory (CT-ML wiki): Distribution Monad · Giry Monad
What it is
A finite, weighted sum of point masses, (equal weights when weights is
nothing): an element of the finite distribution monad used as an
approximation of a continuous distribution. It is an approximation by construction
(isexact is false), and two different particle sets can represent the same distribution.
What it cannot do yet
- Pooling.
combineof two sample beliefs needs the density of at least one of them (importance reweighting), andbelief_logdensityis not defined for particles, socombinethrows. A Dirac or the trivial belief still combines with it. - Pooling with evidence. A density or
AbstractPotentialBeliefcan reweight particles viabelief_logdensity; two clouds alone still have no product rule. RatioFactor emits such a potential, and Mycelium warns on severe ESS collapse (ratio, LP-016). - Convergence checks.
belief_distancecompares weights and corresponding particles; different particle counts giveInf. It does not compare distributions independently of the particle ordering. - A point. The diffusion factor’s
assemble_stateignores sample beliefs, because a particle set has no single value to clamp to.
It is the representation a sampling inference would return (DPS, Langevin; Inference Signatures #6). Weighted particles can represent several branches of a multivalued relation; mixtures provide another representation.
Related: Beliefs, Dirac Belief, Gaussian Belief, Inference Signatures