Aleator.Library 0.3.0-preview.1

Aleator.Library

Standard element library of Aleator, the .NET 10 / C# port of the Figaro probabilistic programming library:

• Atomic continuous distributions — Normal, Uniform, Exponential, Gamma, Beta, Dirichlet, InverseGamma, LogNormal, StudentT, Cauchy, Laplace, Weibull, Pareto, ChiSquared, MultivariateNormal (Cholesky), KernelDensity.
• Atomic discrete distributions — Binomial, Poisson, Geometric, NegativeBinomial, Hypergeometric, Multinomial, UniformInt, FromRange, SwitchingFlip, IntSelector (with OneShifter MH proposals).
• Element-parameter (Compound*) forms — CompoundNormal, CompoundExponential, CompoundGamma, CompoundBeta, CompoundUniform, CompoundDirichlet, CompoundBinomial, CompoundPoisson, CompoundSelect, CompoundDist. These take their parameters from other elements, which is what makes a hierarchy a graph edge rather than a hand-rolled Chain: inference, learning discovery and the gradient bridge all read the dependency off Arguments.
• Truncated — any distribution that answers ICumulativeDistribution (Normal, Exponential, Uniform, LogNormal, Gamma, Beta, InverseGamma, StudentT, ChiSquared, Cauchy, Laplace, Weibull, Pareto, and Truncated itself) restricted to an interval and renormalised over it. Drawn by inverting the CDF, so no sample is rejected and a window in the far tail is sampled as readily as one at the mode.
• Censored — a distribution clipped to an interval, the mass beyond a bound moved onto it (the Tobit model of a saturating measurement): the inner density inside, the censored mass at a bound, so an observation recorded at a detection limit is scored as "at or beyond it". Mixture — a finite mixture with fixed weights as one atomic distribution, with a density, a CDF when the components have one, and draws.
• Deterministic combinators — CPD / RichCpd pattern matching, Inject, tuples, boolean/arithmetic operators.
• Open-universe collections — Container, Process, FixedSizeArray, MakeArray, VariableSizeArray, MakeList.
• Decisions (Decision, policies), references, nonparametrics (Dirichlet-process mixture, CRP, IBP), and the Aleator.Language.Alea static factory facade.

using static Aleator.Language.Alea;

var bias = Beta(2, 5);
var coin = Flip(bias);

// A hierarchy: each level takes the one above it as a parameter.
var mu = Normal(0, 10);
var theta = Normal(mu, 1.0);
var y = Normal(theta, 1.0);

// A scale that cannot be negative, sampled exactly rather than by rejection.
var sigma = TruncatedNormal(0, 3, lower: 0);

Zero third-party dependencies. Pair with Aleator.Algorithms for inference.

Docs and examples: github.com/deecalov/Aleator.