Work

cusna: GPU-accelerated SAOM simulator and amortized neural posterior estimation experiment suite (npe-paper-v1)

Year: 2026
Type: software
Abstract: Companion code for the manuscript "Amortized neural Bayesian inference for stochastic actor-oriented models" (submitted to Social Networks). The archive contains: (i) a CUDA/CuPy forward simulator for stochastic actor-oriented models (SAOMs), validated against RSiena 1.6.6 across six scenarios (moment targets, simulated statistic distributions, and complete method-of-moments estimates; see Appendix B of the manuscript); (ii) a GPU method-of-moments estimator mirroring siena07; (iii) the complete amortized NPE experiment suite (prototype/npe): corpus generation with content-addressed caches, neural spline flow training via the sbi toolkit, simulation-based calibration, posterior predictive checks in the style of sienaGOF, multi-group amortization, and experiment scripts exp01–exp10 that regenerate every number and figure in the paper. Datasets: the s50 excerpt is distributed with the RSiena package; Glasgow and van de Bunt data are fetched by bench/fetch_data.py. Setup notes (Windows, CUDA 12.x, Python 3.12, R 4.6) are in README.md.
Language: English
Open Access status: green
OpenAlex ID: W7167246712
Cites:
Cited by:
Related to: 10