iscc.inference.ABC ¶

ABC(prior, simulate, n_workers=1, seed=0)

Approximate Bayesian Computation engine for fitting model parameters.

Model-agnostic and dependency-light (numpy + scikit-learn + scipy — no JAX, no R). Given a Prior and a simulate(theta) -> summary_vector callback, it builds a reference table, keeps the accept_frac closest draws to the observed summary, and returns a Posterior — with optional Beaumont et al. (2002) local-linear regression adjustment and a Random-Forest point estimate (the Python analogue of CINner's ABC-rf). The simulate callback is supplied by the caller (for the tumour model, the iscc.inference package wires one up). Call run with the observed summary vector to obtain the posterior.

Parameters:

Name	Type	Description	Default
`prior`	`Prior`	Product prior over the named parameters to infer (see `Prior`).	required
`simulate`	`callable`	`dict(name -> value) -> summary vector`; a single forward simulation of the model.	required
`n_workers`	`int`	Number of processes used to parallelise the simulations (default 1 = serial).	`1`
`seed`	`int`	Seed for the engine's random generator (default 0).	`0`

Methods:

Name	Description
`reference_table`	Draw `n_samples` priors and simulate them -> `(theta, summaries)` (failures dropped).
`run`	Run ABC against `observed` -> `Posterior`.

reference_table ¶

reference_table(n_samples)

Draw n_samples priors and simulate them -> (theta, summaries) (failures dropped).

run ¶

run(observed, n_samples=1000, accept_frac=0.1, adjust=True, rf=True, project='rf', reference=None)

Run ABC against observed -> Posterior.

reference optionally reuses a precomputed (theta, summaries) table (so the same sims can serve several observations). accept_frac sets the rejection tolerance.

project="rf" enables semi-automatic ABC (Fearnhead & Prangle 2012): a random forest is fit to map summaries -> parameters, and the rejection/regression step runs in that low-dimensional predicted-parameter space. Each projected coordinate targets one parameter on a comparable scale, so a single clean summary can no longer dominate the distance and starve the others (the failure mode of raw standardised-Euclidean matching when parameters are identified by different summaries). project=None uses the raw standardised summaries. The reference table is projected through the forest's out-of-bag predictions to avoid the train-on / match-on double-dip.