Distributions#

Beliefs, priors and parameter posteriors are distribution objects: immutable equinox modules that hold their parameters as arrays. The conditional families give a child distribution for every configuration of its parents, either by table or by a small network. Exponential-family members expose their natural and mean parameters through one shared contract.

Base types#

Distribution

Abstract base for every belief, message payload and preference object.

ExponentialFamily

A Distribution in the exponential family, with a natural-parameter view.

Learnable

Structural capability protocol (see module docstring), satisfied by method presence.

Conditional

Abstract base for conditional families: Markov kernels P(child | parents) with discrete parents.

Distributions over states#

Categorical

A categorical distribution over num_levels outcomes, batched over leading axes.

JointCategorical

Joint distribution over event_ndims discrete axes.

Gaussian

A multivariate Gaussian with full covariance, batched over leading axes.

MixtureGaussian

Gaussian mixture belief.

TruncatedGaussian

A diagonal Gaussian N(loc, diag(scale²)) restricted to [low, high] (see module).

Dirichlet

A Dirichlet distribution over the num_levels simplex, batched over leading axes.

Conditional families#

ConditionalCategorical

Tabular CPD P(child | parents): one simplex per parent configuration.

ConditionalGaussian

Parent-indexed stack of Gaussians: P(x | parents) with discrete parents.

ConditionalDirichlet

Independent Dirichlet priors over a CPD, one per parent configuration.

MLPCategorical

MLP over concatenated parents → a factored categorical child.

MLPConditional

MLP over concatenated parents → diagonal-Gaussian child (see module docstring).

MLPPredictor

MLP over concatenated parents → a raw output vector, with no density of its own.

PredictorHead

An MLPPredictor plus the name of the loss that reads its output.

LearnedVector

A trainable vector behind an activation: a parametrized constant, with no parents.

Parameter priors and further families#

These are exported by hybrid_aif.distributions only.

MatrixNormal

W ~ MN(M, U, V) with shared column covariance; row structure external.

InverseGamma

Independent inverse-Gamma posteriors over per-dimension noise variances.

InverseWishart

Inverse-Wishart posterior over a full noise covariance.

LowRankMixture

Gaussian mixture whose component covariances are diagonal plus low rank, S_i = diag(D_i) + U_i U_iᵀ, with U_i of r columns in a p-dimensional space.

GRUCell

A GRU carry update (deter, *inputs) -> deter': a pure map, not a conditional.