AGNet — Learning to Place Guards by Reinforcement

Paper: Learning to Place Guards by Reinforcement: A Geo-Free Neural Policy for the Vertex-Guard Art Gallery Problem Domagoj Ševerdija, Jurica Maltar, Nathan Chappel, Domagoj Matijević

This repository contains the code and data for the paper. We study whether an LSTM pointer-network policy trained by reinforcement learning on the vertex-guard Art Gallery Problem (AGPVG) learns a representation that encodes the underlying geometry — independent of what its decoder expresses. The key findings:

• A pointer-network policy trained geo-free (only vertex coordinates at inference, no visibility oracle) places guards at near-greedy cardinality but leaves a tail of under-covered polygons.
• Freezing the encoder and probing it with a small single-shot SetPredictor classifier closes most of the feasibility gap, in and out of distribution (up to 5× the training range), cutting infeasible polygons by roughly an order of magnitude.
• A no-encoder ablation and a linear probe (ROC-AUC = 0.842) confirm that the representation — not the probe's own capacity — carries the guard-relevant geometry.

---

Repository structure

AGNet/
├── po_agp.py                   # PO/BT policy training (main entry point)
├── train_set_predictor.py      # SetPredictor probe training
├── eval_set_predictor.py       # Evaluation: threshold sweep + iterative refinement
├── eval_checkpoint.py          # Evaluate policy checkpoint (CGAL visibility)
├── models.py                   # PointerNet encoder/decoder architecture
├── set_predictor.py            # SetPredictor architecture
├── rewards.py                  # Coverage + guard-count reward definitions
├── dataset.py                  # AGPVG dataset loader
├── greedy_agp.py               # Classical greedy baseline
├── tools/
│   └── build_ls_trajectories.py  # Generate LS-editor training targets
├── configs/                    # JSON configs for all experiments
│   ├── po_agp_lstm.json
│   └── set_predictor_train_standard.json
└── data/                       # Precomputed LS trajectories and greedy baselines
    ├── ls_trajectories_train.pkl
    ├── ls_trajectories_dev.pkl
    ├── ls_trajectories_test.pkl
    └── ...

---

Installation

Requirements: Python 3.10+, PyTorch 2.0+, CGAL (for exact visibility at evaluation).

git clone git@github.com:dseverdi/AGNet.git
cd AGNet
python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

CGAL / scikit-geometry

Exact visibility (used only at evaluation, not training) requires scikit-geometry:

# Recommended: via conda
conda install -c conda-forge scikit-geometry

# Alternative: build from source (see scikit-geometry docs)

---

Dataset (AGPVG)

The polygon instances come from the Art Gallery Problem Vertex-Guard (AGPVG) benchmark:

Couto et al., Art Gallery Problem Vertex Guard instances https://www.ic.unicamp.br/~cid/Problem-instances/Art-Gallery/AGPVG/

Download and extract the dataset, then configure the path:

cp .env.example .env
# Edit .env and set DATASET_PATH to the extracted AGPVG directory

The data/ directory in this repo already contains precomputed LS-editor trajectories (training targets for the SetPredictor) and greedy baseline results — you do not need to regenerate these to run evaluation.

---

Quick start: run the trained probe

Download the pretrained checkpoints from the GitHub Releases page:

File	Description
po_agp_best_greedy.pt	PO/BT LSTM pointer-network policy
set_predictor_best.pt	SetPredictor probe (standard τ=0.99 targets)

Place them at checkpoints/v3/po_agp/lstm_bt/po_agp_best_greedy.pt and checkpoints/set_predictor/standard/set_predictor_best.pt, then:

# Threshold sweep on the dev_test split (reproduces Table 3 in the paper)
python eval_set_predictor.py \
    --checkpoint checkpoints/set_predictor/standard/set_predictor_best.pt \
    --split dev_test

# Full Pareto sweep across thresholds and inference passes (Tables 4–5)
python eval_set_predictor.py \
    --checkpoint checkpoints/set_predictor/standard/set_predictor_best.pt \
    --split test \
    --iter-passes-sweep 1 2 3 5

---

Reproducing paper results from scratch

1. Train the PO/BT policy

python po_agp.py configs/po_agp_lstm.json
# Checkpoint saved to checkpoints/v3/po_agp/lstm_bt/po_agp_best_greedy.pt

2. Build LS-editor training targets (optional — precomputed data included)

The data/ls_trajectories_*.pkl files are already in the repo. To regenerate from scratch:

python tools/build_ls_trajectories.py \
    --split train \
    --checkpoint checkpoints/v3/po_agp/lstm_bt/po_agp_best_greedy.pt \
    --out data/ls_trajectories_train.pkl

3. Train the SetPredictor probe

# Four seeds as used in the paper: 1234, 11, 22, 33
for SEED in 1234 11 22 33; do
    python train_set_predictor.py configs/set_predictor_train_standard.json \
        --seed $SEED \
        --out-dir checkpoints/set_predictor/standard_seed${SEED}
done

4. Evaluate

python eval_set_predictor.py \
    --checkpoint checkpoints/set_predictor/standard_seed1234/set_predictor_best.pt \
    --split dev_test

---

Configuration

All experiments are driven by JSON configs in configs/. Key files:

Config	Purpose
configs/po_agp_lstm.json	PO/BT LSTM policy training
configs/set_predictor_train_standard.json	SetPredictor probe (τ=0.99 targets)

---

Citation

@article{severdija2026agnet,
  title   = {Learning to Place Guards by Reinforcement: A Geo-Free Neural Policy
             for the Vertex-Guard Art Gallery Problem},
  author  = {Ševerdija, Domagoj and Maltar, Jurica and Chappel, Nathan
             and Matijević, Domagoj},
  journal = {Preprint},
  year    = {2026}
}

---

License

Code: MIT. Dataset: see AGPVG page for terms.