Generation-free ABC sampler updates proposals as simulations finish
Original title: Asynchronous, Generation-Free ABC: Streaming AMIS Reweighting for Heterogeneous Simulator Workloads on HPC
On a Cellular Potts tissue simulator it finished 2.1–2.4 times as many evaluations as pyABC in one hour.
Researchers describe an asynchronous, generation-free approximate Bayesian computation sampler for heterogeneous simulator runtimes on HPC. It updates the proposal as results arrive, reuses the full particle history, and reweights against a reconstructed mixture of proposals. Experiments report better-calibrated posteriors than a pyABC baseline with the same kernel and bandwidth schedule, and predicted synchronization slowdowns of 1.2x to 400x matched measurements.
- Workers receive a new candidate as soon as a simulation result arrives.
- The best 100 evaluations had a discrepancy threshold 4.1 times smaller than pyABC.
- Starting bandwidth from prior-predictive discrepancies yielded a more concentrated two-parameter posterior.
- Consistency is shown for the smooth-ABC posterior when the mixture approximation is asymptotically exact.
From the original article
1 min read at arxiv.orgApproximate Bayesian computation (ABC) enables Bayesian inference for models with intractable likelihoods by estimating parameter posterior distributions from simulated data. Sequential ABC methods improve sampling efficiency by iteratively adapting the proposal distribution toward regions of high posterior probability.
By Julian Herold · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.