Trust-DVine-4D: Simulation-Based Inference Benchmark (4D)

This dataset contains simulation-based inference results for the 4-dimensional DVine problem from the "Averting A Crisis in Simulation-Based Inference" paper.

Dataset Description

The DVine (Dependency Vine) problem evaluates Bayesian inference methods on a 4-dimensional parameter space using a vine copula-based simulator. This dataset contains:

Reference Data (vine_reference/)

• coverage-vine-posterior.npy - Coverage statistics for the vine posterior
• coverage-vine-mst.npy - Coverage statistics for the MST (Minimum Spanning Tree) vine
• sbc-vine-posterior.npy - Simulation-based calibration results for vine posterior
• sbc-vine-mst.npy - SBC results for MST vine

Simulation Results (output/{sample_size}/)

Results organized by number of simulations (sample sizes: 1024, 2048, 4096, 8192, 16384, 32768, 65536, 131072).

Each sample size directory contains results from multiple methods (without-regularization/):

• MLP methods: mlp-{00-04}, mlp-bagging-{00-04}, mlp-static-{00-04}
• Flow-SBI methods: flow-sbi-{00-04}, flow-sbi-bagging-{00-04}, flow-sbi-static-{00-04}

File Types

• coverage.npy - Coverage diagnostic arrays
• sbc.npy - Simulation-based calibration arrays
• diagnostic.npy - Additional diagnostic metrics
• losses-*.npy - Training/validation/test losses
• test-loss-functionals.npy - Test loss functionals
• posterior.pkl - Serialized posterior samples

Citation

If you use this dataset, please cite:

@article{terral2022averting,
  title={Averting A Crisis in Simulation-Based Inference},
  author={Terral, Thomas and others},
  journal={arXiv preprint arXiv:2110.06581},
  year={2022}
}

Related Datasets

• k-kiirikki/trust-dvine-6d - 6-dimensional DVine results