Conway Life: Local MLP Predictors, Rollouts, and Dynamics

This repo trains and evaluates MLPs that approximate Conway’s Game of Life.
We focus on (1) single-step patch prediction for 3×3/5×5/7×7 neighborhoods,
(2) autoregressive rollouts with/without calibration, and (3) a dynamics baseline that maps a full 3×3 patch to the next-step center cell.

What’s here (kept files)

• data.py / game_of_life.py: dataset generation (toroidal wrap), GoL simulator.
• models.py: MLPs for patch classifiers.
• train_weighted.py: trains 3×3, 5×5, 7×7 patch models on the general dataset (p=0.5, wrap) with class weighting w_pos = min(5, sqrt(r)). Saves to checkpoints_weighted/.
• train_weighted_regimes.py: train/evaluate per-regime/density if needed.
• eval_patch_models.py: evaluate existing checkpoints on test split; can write CSV.
• batch_autoregressive_eval.py: batched rollouts (default alignment current) with options:
  • --model_variant {general,match} selects general ckpt vs regime/density-matched ckpt.
  • --calibrated enables temp/threshold scan + optional adaptive thresholding.
  • Outputs snapshots/curves/CSV to rollout_batch_viz/<calibrated|noncalibrated>/<general|matched>/.
• train_dynamics_patch3.py: 9→1 dynamics model (full 3×3 → next center), with early stopping and board-level rollouts; checkpoints in checkpoints_dynamics/.
• rollout_batch_viz/: current rollout figures and CSVs (calibrated & noncalibrated, general & matched).
• legacy_viz/: older confusion-matrix figures.
• checkpoints_weighted/, checkpoints_dynamics/, data/: trained models and datasets (kept).

Typical commands

Train general weighted models (wrap data, patch 3/5/7):

python train_weighted.py

Evaluate a checkpoint on the test split:

python eval_patch_models.py --ckpt checkpoints_weighted/best_model_patch5_life_patches_epoch15.pth

Run non-calibrated rollouts (general models, alignment=current):

python batch_autoregressive_eval.py \
  --patch_sizes 3 5 7 \
  --regimes early mid late \
  --densities 0.2 0.4 0.6 \
  --seeds 0 \
  --steps 50 \
  --model_variant general
# results → rollout_batch_viz/noncalibrated/general/

Run calibrated rollouts (temperature + threshold scan, adaptive thresholding enabled):

python batch_autoregressive_eval.py \
  --patch_sizes 3 5 7 \
  --regimes early mid late \
  --densities 0.2 0.4 0.6 \
  --seeds 0 \
  --steps 50 \
  --model_variant general \
  --calibrated
# results → rollout_batch_viz/calibrated/general/

Matched models (if per-regime/density checkpoints are available) use --model_variant match and save under rollout_batch_viz/.../matched/.

Train the dynamics 3×3→next-center model and test board rollouts:

python train_dynamics_patch3.py --generate --max_epochs 10 --patience 2

Data & checkpoints

• Default dataset: data/life_patches.npz (wrap extraction, includes 3/5/7 patches).
• General checkpoints: checkpoints_weighted/best_model_patch{3,5,7}_*.pth.
• Dynamics checkpoint: checkpoints_dynamics/best_model_dynamics_patch3.pth.

Outputs

• Rollouts: density curves + accuracy/F1 curves + board snapshots per run, plus CSV with avg_acc/avg_f1/precision/recall and chosen temp/threshold (for calibrated runs).
• Dynamics script: training/val logs, test metrics, and short board rollouts printed to stdout.

Notes

• Everything uses toroidal wrap to avoid edge artifacts.
• Class imbalance handled via positive class weight min(5, sqrt(r)).
• Calibration (when enabled) scans temps {1, 1.5, 2} and thresholds {0.6, 0.7} on a short horizon, then runs full rollout with optional adaptive threshold tied to predicted density.

Progress so far

• Early phase: evaluated patch-level predictors across densities/regimes; recall was low, so we adopted the weighted loss (V1) based on F1 trade-off.
• Autoregressive phase: rolled out single-step models; severe “explosion to all-alive” prompted calibration (temp + threshold + optional adaptive tweak). Calibration lifts accuracy but F1 remains modest and drift persists.
• Dynamics probe: trained a 3×3→next-center model using wrapped data; achieved perfect test and board-level rollouts (deterministic mapping is learnable).
• Next steps: use the autoregressive setup to check whether the model reproduces stable GoL motifs (still lifes, oscillators, gliders) or collapses to trivial attractors; refine calibration/thresholding if needed.