Design

foerster is inference built from one primitive of spindel: the savepoint. This guide explains how, for readers who want to write an algorithm or understand the ones there are. The savepoint itself — its laws, sessions, worlds — is described in spindel's savepoints design.

Sites are savepoints

A spindel savepoint is a point a program publishes: the rest of the computation, offered to a handler, which may resume it (once), fork it into new worlds (any number of times), copy it (consuming it), or abandon it. Worlds are copy-on-write, so forking is cheap and a fork continues exactly where the savepoint was, with the world as it was.

sample, observe and factor publish savepoints when the world they run in handles their site, and simulate forward otherwise:

Site	Payload	A policy resumes it with
:inference/choose (sample)	{:dist d :options …}	a draw from d, a kept value, a constraint, or a proposal's draw — having added log d(v) − log q(v) to the weight
:inference/choose (observe)	{:dist d :observed? true :value y}	y, having added log d(y)
:inference/factor	{:log-weight w}	nil, having added w

Every inference algorithm is a handler over these sites, and they share the same programs. This is Anglican's design — a model's sample and observe are checkpoints of a continuation-passing interpreter — with the continuation now a forkable world.

Traces and policies

A trace (spindel.trace, foerster.trace) records one run: every site's address, value and notes. The notes of an entry are :dist, :log-prob — the log density of every site, sampled ones included, not only observations — :log-proposal when the value was drawn from a proposal, and flags: :observed?, :constrained?, :kept?, :symmetric?, :factor?, :intervened?. The importance weight accumulates at [:inference :log-weight] in the world.

A policy (trace/policy) decides sites: constraints, kept values from an earlier trace, draws from custom proposals, interventions, counterfactual noise, and the prior for everything else. trace/run and trace/replay run a program under a policy.

The algorithms as handlers

• SMC (foerster.smc) forks the first site into N particle worlds (in canonical worlds: copies the root at a start site, so every particle runs the whole model), runs each until it parks at an observation, and when all have arrived resamples if the effective sample size is below :resample-threshold·N: the chosen ancestors' parked savepoints are forked (or copied, in canonical worlds) into the next generation, the others abandoned. No coordinator and no barrier thread: the barrier is the arrival of the last particle. Streaming SMC parks particles at :stream sites until a value is pushed. In fresh worlds the session is closed before the measure is delivered, so a caller that stops its executor on the result cuts no cleanup short; in canonical worlds every world is discarded first (below).

• Importance sampling is the same with :resample-threshold 0: particles still park at each observation, but the population is never resampled.

• A PInferenceKernel (kernel-infer) runs as SMC whose latent sites take the value the kernel's step gives, its :log-weight-delta added to the weight; the prior kernel draws from the prior.

• The particle cascade (foerster.cascade) has no barrier: each savepoint arriving at an observation is decided on its own, against the running mean weight of those that arrived there so far, itself included — abandoned, continued, or forked into several children (Paige et al. 2014, Eq. 14).

• Steering (foerster.steer) is SMC over a program whose steps are random without a sample site and whose value estimates are barrier factors: twisted SMC, with the telescoping correction as the last factor.

• Conditional SMC keeps one particle on a retained trajectory: particle Gibbs, and with ancestor sampling (PGAS) the retained particle redraws its past, each candidate's future scored by forking its parked savepoint under a scoring handler.

• PIMH proposes a whole SMC sweep and accepts it on the ratio of evidence estimates; IPMCMC runs several SMC and conditional-SMC nodes and exchanges which of them are conditional by Gibbs updates on their evidence.

• Metropolis–Hastings is replay plus an accept step (trace/mh-step, mh-chain). A move selects a set of target sites (one for single-site MH, a block for block Gibbs), replays from the earliest, proposes at the targets, and keeps every other site's value rescored under its distribution as it is now:

  log a = [log p(new) − log p(old)]
        + [log q(old | new) − log q(new | old)]
        + [log s(targets | new) − log s(targets | old)]
  

  where q(new | old) is what the move drew afresh, q(old | new) what the reverse move would have to draw of the old trace, and s the probability of selecting the targets, which changes when the move changed how many sites there are. Single-site MH proposes from the prior; random-walk MH perturbs continuous sites with a normal step and is symmetric; block Gibbs selects blocks of sites by a selector and moves each with its own kernel. A kept value that fell out of its site's support is drawn again; a move whose reverse would keep the new value (overlapping supports) is refused. Limits: the reverse move is scored under the prior, so a custom proposal must be the prior or symmetric, and a block's membership must not depend on the move. Chain kernels compose: k/cycle runs several in turn, k/mixture one picked at random per iteration; both are kernels again and nest.

• HMC-within-Gibbs moves block sites along the gradient of the block's density — plus one other latent site by single-site MH per iteration — and accepts on the full trace's log joint, so an incomplete block density costs mixing, not correctness. Under the particle methods a block site is drawn from its :sample capability and weighted by its density over the density of the draw (:sample-log-density, required there).

• Involutive MCMC (foerster.involutive): an auxiliary draw and an involution on (choices, aux) with its log Jacobian.

• BBVI learns a mean-field q by stochastic gradient ascent on the ELBO with score-function gradients (foerster.gradient) and control variates; its samples are importance samples from q.

• The generative function interface (foerster.gfi) — simulate, generate, assess, update, regenerate — is Gen's, on the same traces.

• Counterfactuals (foerster.counterfactual) read every site as a structural equation of exogenous noise (foerster.mechanism): abduction infers the noise, action intervenes, prediction replays.

Canonical worlds

With :world-policy :fork the model runs in a frozen fork of the caller's world, the root, and its session's worlds join the root's scope; every particle runs the whole model, is a copy of a world (grades checked, budgets split), and every world is discarded before the result is delivered. See worlds. Markov chains run in fresh worlds only.

Not implemented yet

• :parents in a site's payload: which earlier choices a site read — which no address scheme can recover — making a trace a graph (variables, edges, observed mask, values), the training datum of graphically structured amortized inference. (:proposal, a guide distribution the program computes, is implemented: a fresh draw comes from it and the weight takes log p − log q.)
• Markov chains in canonical worlds.
• Resample-move inside conditional SMC (particle Gibbs, PGAS), refused today: keeping the retained particle unmoved while the others move leaves the conditional target off (an exact enumeration of one sweep puts its stationary law off by 2·10⁻⁴ to 3·10⁻³ in total variation). The exact construction treats the retained path as the state after each barrier's move and draws its state before the move backwards through the move's reversal (for a reversible MH kernel, the kernel itself) — exact to 10⁻¹⁵ in the same enumeration. It needs the retained path drawn backwards through every barrier's move before a sweep, and slot 0 switched from its pre-move to its post-move values at each barrier.

What else the same programs admit

Method	How
reweighted wake-sleep, inference compilation	per-site :log-prob and :log-proposal in the trace; gradients by the embedding
nested inference	an inner world with its own handler table; nesting is by world