AutonomousOED

This repository contains the scripts used to reproduce the performance benchmarks and (with access to the large CSZ dataset) the greedy sensor-selection application described in paper: "Sensor Placement for Tsunami Early Warning via Large-Scale Bayesian Optimal Experimental Design."

Repository layout

• create_test_mat.py: create a synthetic 2D-chunked HDF5 matrix for scaling benchmarks.
• single_gpu_benchmark.py: single-GPU benchmark for naive vs Schur-update formulations.
• scaling_benchmark.py: multi-GPU benchmark with overlapped POSIX I/O + compute.
• run_scaling.sh: Slurm wrapper to run both strong + weak scaling.
• select_sensors.py: MPI greedy selection on an HDF5-backed CSZ (K) matrix (real dataset required).
• oed_hist.py: generate random-configuration baselines and (optionally) plot the histogram.
• d_opt_certificate.py / scripts/certify_d_optimality.py: a-posteriori optimality certificates for existing regular D-optimal designs.
• tests/test_certify_d_optimality.py: small CPU synthetic certificate tests.

Installation

Create an environment and install dependencies:

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

For parallel file systems, it is recommended to disable HDF5 file locking:

export HDF5_USE_FILE_LOCKING=FALSE

Task 1: Generate dummy data (for scaling benchmarks)

python create_test_mat.py --file dummy_data.h5 --total_candidates 600 --Nt 420 --precision double

Task 2: Single-GPU benchmark (Figure 2)

python single_gpu_benchmark.py --max_budget 400 --Nt 420 --file ablation_results_420_single.csv

The output CSV contains peak memory and time-per-iteration for:

• Naive formulation (out-of-place / in-place)
• Schur-update formulation (out-of-place / strict in-place)

Task 3: Multi-GPU scaling benchmarks (Figure 3)

This script assumes Slurm and srun. Provide the HDF5 file and the maximum rank count to test.

bash run_scaling.sh --data dummy_data.h5 --max-ranks 512

Outputs are written under scaling_results/ by default (both strong and weak scaling CSVs).

Task 4: Cascadia application (Figure 5, real dataset required)

To download the dataset, visit the HuggingFace repo and follow the instructions there.

With the real HDF5 (K) matrix available on a parallel filesystem, run the greedy selection via MPI. Example:

srun -n 16 python select_sensors.py cascadia_K_matrix.h5 175 \
  --r_sq 1.0 \
  --checkpoint_file cascadia_optimal_results.txt

Task 5: Random baseline + histogram (Figure 4)

Generate random baseline scores (MPI-capable; produces a single-column checkpoint file). This run skips plotting:

python oed_hist.py \
  --h5_store_K cascadia_K_matrix.h5 \
  --total_samples 100 \
  --budget 175 \
  --checkpoint_file random_results.txt \
  --no_plot

Plot histogram comparing random baselines to the greedy optimum:

python oed_hist.py \
  --plot_only \
  --checkpoint_file random_results.txt \
  --optimal_file cascadia_optimal_results.txt \
  --budget 175

Task 6: A-posteriori D-optimality certificates

Given an existing ordered regular D-optimal selection (e.g. from Task 4) and the same HDF5 signal covariance, certify how close the design is to the globally optimal (k)-sensor design without enumerating all designs.

Noise convention matches select_sensors.py: the HDF5 file stores the signal covariance (C), and the working data covariance is the isotropic model

[ K = r^2 C + I. ]

Reported values use the normalized expected information gain (f(S)=0.5\log\det(K_S)) (since (R=I)). This is not the unnormalized logdet column written in greedy checkpoint files, and it is not a posterior trace / A-optimality metric.

Cheap certificate (default checkpoints (0) and (k)):

python scripts/certify_d_optimality.py cascadia_K_matrix.h5 \
  --selection cascadia_optimal_results.txt \
  --r_sq 1.0 \
  --output_prefix certificates/cascadia_k175

More expensive multi-checkpoint analysis (rescans remaining candidates at each prefix; dense schedules can approach the cost of re-running greedy):

python scripts/certify_d_optimality.py cascadia_K_matrix.h5 \
  --selection cascadia_optimal_results.txt \
  --r_sq 1.0 \
  --checkpoint_stride 25 \
  --output_prefix certificates/cascadia_stride25

Outputs:

• <prefix>_checkpoints.csv — per-checkpoint (f(S)), top-(k) remaining marginal sum, (U(S)), and running (U_{\mathrm{best}}).
• <prefix>_summary.json — machine-readable summary including certified ratio (f(S_{\mathrm{final}})/U_{\mathrm{best}}) and absolute gap bound.

Small CPU synthetic tests (no production dataset required):

python -m unittest tests.test_certify_d_optimality -v

Scope: regular D-optimality with independent isotropic sensor-noise blocks only. Goal-oriented D-optimality and cross-sensor-correlated noise are rejected. --r_sq is required and must match the value used to produce the selection.

Notes for artifact packaging

• Large datasets (e.g., the CSZ (K) matrix) are intentionally not included in this repository.
• All scripts accept file paths via CLI arguments.