BEACON

Companion code for BEACON: A Bayesian Optimization Inspired Strategy for Efficient Novelty Search. BEACON has been published at the Digital Discovery Journal.

BEACON uses Bayesian optimization ideas to search for novel behaviors in expensive black-box systems. This repository is organized as a paper companion: experiment scripts remain close to the manuscript studies, while shared inputs, saved outputs, trained weights, and plotting data live in predictable top-level folders.

The animations below illustrate the sampling progress of Random Sampling (RS) and BEACON on the synthetic multi-output problem considered in the paper. BEACON efficiently explores the entire outcome space, achieving broad and uniform coverage throughout the process. In contrast, RS struggles to escape the central outcome region, where the majority of the function's outcomes are concentrated, resulting in substantially poorer exploration. Quantitative reachability results are presented in Figure 3 of the paper.

Repository Map

Path	Contents
beacon/	Shared Python helpers, including repo-relative paths and efficient Thompson sampling.
experiments/continuous/single-outcome/	Ackley, Rosenbrock, Styblinski-Tang, and noisy Ackley scripts.
experiments/continuous/multi-outcome/	Synthetic multi-outcome and maze scripts.
experiments/discrete/single-outcome-material/	Material and molecule single-outcome scripts.
experiments/discrete/multi-outcome/	Oil sorbent, joint gas uptake, and MNIST multi-outcome scripts.
figures/scripts/	Plotting scripts used to recreate paper-style figures from saved data.
data/materials/	Tracked tabular and pickle inputs used by material-discovery scripts.
results/continuous-single-outcome/	Tracked saved tensor outputs for continuous single-outcome plots.
results/plot-data/	Saved .mat files and image assets that plotting scripts read.
results/generated/	Local tensor outputs written by experiment scripts. Ignored by git.
models/mnist/	Tracked CNN/VAE weights used by the MNIST script.
tests/	Lightweight repository checks; these do not run full experiments.

results/plot-data/ means "data used by plotting scripts," not generated figures. Generated figures should be treated as local outputs, and figures/output/ is ignored by git for that purpose.

Installation

Use a repo-local virtual environment:

python3 -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -r requirements.txt

On machines where the default python3 is newer than the scientific stack supports, Python 3.11 is a good target. For example, with uv installed:

uv venv --python 3.11 --seed --clear .venv
source .venv/bin/activate
python -m pip install -r requirements.txt

Some experiments are computationally heavy or require optional environments. In particular, the maze script uses gymnasium-robotics, and the MNIST script can download MNIST data into data/mnist/.

Running Scripts

Run scripts from the repository root so their repo-relative paths resolve consistently:

python experiments/continuous/single-outcome/ackley-beacon.py
python experiments/discrete/single-outcome-material/logd-beacon.py
python figures/scripts/plot-synthetic.py

The optimization scripts are intended to reproduce paper experiments and may take a long time. When an experiment finishes, it writes generated tensors to results/generated/<domain>/<task>/<script-name>/.

Plotting scripts read from results/continuous-single-outcome/ or results/plot-data/ and write generated figures into figures/output/.

Shared Code

beacon/thompson_sampling.py implements the efficient Thompson sampling strategy used by BEACON, based on Wilson et al., Efficiently Sampling Functions from Gaussian Process Posteriors.

beacon/paths.py defines repo-root-relative paths used by scripts:

from beacon.paths import MATERIALS_DIR, PLOT_DATA_DIR, GENERATED_RESULTS_DIR, FIGURE_OUTPUTS_DIR

Checks

The lightweight tests check syntax, importability of shared helpers, artifact placement, stale absolute paths, naming conventions, and removal of tracked bytecode:

source .venv/bin/activate
python -B -m pytest

These checks deliberately do not run the full BEACON experiments, MNIST downloads, or maze simulations.

Before merging a cleanup branch back to main, confirm:

• python -B -m pytest passes in .venv.
• Artifact checksums match before and after structural moves.
• README commands match the actual paths.
• No active code imports retired src.* paths or references stale absolute paths.
• git diff --name-status --find-renames repo-cleanup-companion...repo-structure-redesign is readable.

Extension To High-Dimensional Problems

The high-dimensional extension, TR-BEACON, is described in the paper and maintained in the TR-BEACON repository.

Citation

If you use this code in your research, please cite:

@article{tang2026beacon,
  title={BEACON: A Bayesian Optimization Inspired Strategy for Efficient Novelty Search},
  author={Tang, Wei-Ting and Chakrabarty, Ankush and Paulson, Joel},
  journal={Digital Discovery},
  year={2026},
  publisher={Royal Society of Chemistry}
}