Aleator.AutoDiff 0.3.0-preview.1

Aleator.AutoDiff

Gradient-based inference stack of Aleator — capabilities that go beyond the original Scala Figaro:

• Reverse-mode autodiff tape (Tape, Var, AdMath) with stable log-sigmoid / log-sum-exp primitives, unconstraining transforms (exp / interval / stick-breaking), and differentiable log-densities.
• JointModel builder + UniverseJointBuilder bridge from Aleator elements to a differentiable joint target.
• Samplers — HamiltonianMonteCarlo, NoUTurnSampler (dual averaging, mass-matrix adaptation).
• Variational inference — MeanFieldAdvi, FullRankAdvi, SteinVariationalGradientDescent.
• Gaussian processes — composable kernels (RBF, Matérn, linear, …), marginal-likelihood hyperparameter learning through reverse-mode Cholesky, FITC sparse approximation.

var model = new JointModel();
var mu = model.Real("mu");
var target = model.Build(s =>
{
    var logp = LogDensities.Normal(s[mu], 0.0, 10.0);
    foreach (var y in data)
        logp += LogDensities.Normal(s.Const(y), s[mu], 1.0);
    return logp;
});
var nuts = new NoUTurnSampler(target, random: new DefaultRandomSource(42));
var draws = nuts.Sample(numSamples: 2000, warmup: 1000);

HMC and NUTS require a finite initial target value and gradient. Non-finite proposals are rejected as divergences; inspect draws.Divergences after sampling. UniverseJointBuilder rejects element evidence because observations, conditions, and constraints must be represented explicitly in its differentiable likelihood callback.

Public result metadata is exposed through read-only snapshots: kernel hyperparameters, GP optimisation history, and variational means, histories, Cholesky rows, and particle rows cannot be changed through an IList cast. UniverseJointBuilder.Latents is instead a live read-only view, so a retained view reflects subsequent Add calls. For compatibility, PosteriorSamples.Draws keeps its IReadOnlyList<double[]> signature; because each row is still a mutable array, every property access returns a fresh deep defensive copy. Cache that returned value when traversing the same draws repeatedly; changing it never affects Mean, Variance, Map, or other sampler diagnostics.

Zero third-party dependencies.

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