Dominik Roth
0932bb353a
- nucon/rl.py: delta_action_scale action space, bool handling (>=0.5), direct sim read/write bypassing HTTP for ~2000fps env throughput; remove uncertainty_abort from training (use penalty-only), larger default batch sizes; fix _read_obs and step for in-process sim - nucon/model.py: optimise _lookup with einsum squared-L2, vectorised rbf kernel; forward_with_uncertainty uses pre-built normalised arrays - nucon/sim.py: _update_reactor_state writes outputs via setattr directly - scripts/train_sac.py: moved from root; full SAC+HER example with kNN-GP sim, delta actions, uncertainty penalty, init_states - scripts/collect_dataset.py: CLI tool to collect dynamics dataset from live game session (--steps, --delta, --out, --merge) - README.md: add Scripts section, reference both scripts in training loop Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
56 lines
2.3 KiB
Python
56 lines
2.3 KiB
Python
"""Collect a dynamics dataset from the running Nucleares game.
|
|
|
|
Play the game normally while this script runs in the background.
|
|
It records state transitions every `time_delta` game-seconds and
|
|
saves them incrementally so nothing is lost if you quit early.
|
|
|
|
Usage:
|
|
python scripts/collect_dataset.py # default settings
|
|
python scripts/collect_dataset.py --steps 2000 --delta 5 # faster sampling
|
|
python scripts/collect_dataset.py --out my_dataset.pkl
|
|
|
|
The saved dataset is a list of (state_before, action_dict, state_after, time_delta)
|
|
tuples compatible with NuconModelLearner.fit_knn() and train_model().
|
|
|
|
Tips for good data:
|
|
- Cover a range of operating states: startup, ramp, steady-state, shutdown.
|
|
- Vary individual rod bank positions, pump speeds, and MSCV setpoints.
|
|
- Collect at least 500 samples for kNN-GP; 5000+ for the NN backend.
|
|
- Merge multiple sessions with NuconModelLearner.merge_datasets().
|
|
"""
|
|
import argparse
|
|
import pickle
|
|
from nucon.model import NuconModelLearner
|
|
|
|
parser = argparse.ArgumentParser()
|
|
parser.add_argument('--steps', type=int, default=1000,
|
|
help='Number of samples to collect (default: 1000)')
|
|
parser.add_argument('--delta', type=float, default=10.0,
|
|
help='Game-seconds between samples (default: 10.0)')
|
|
parser.add_argument('--out', default='reactor_dataset.pkl',
|
|
help='Output path for dataset (default: reactor_dataset.pkl)')
|
|
parser.add_argument('--merge', default=None,
|
|
help='Existing dataset to merge into before saving')
|
|
args = parser.parse_args()
|
|
|
|
learner = NuconModelLearner(
|
|
time_delta=args.delta,
|
|
dataset_path=args.out,
|
|
)
|
|
|
|
if args.merge:
|
|
learner.merge_datasets(args.merge)
|
|
print(f"Merged existing dataset from {args.merge} ({len(learner.dataset)} samples)")
|
|
|
|
print(f"Collecting {args.steps} samples (Δt={args.delta}s each) → {args.out}")
|
|
print("Play the game — vary rod positions, pump speeds, and operating states.")
|
|
print("Press Ctrl-C to stop early; data collected so far will be saved.")
|
|
|
|
try:
|
|
learner.collect_data(num_steps=args.steps)
|
|
except KeyboardInterrupt:
|
|
print("\nInterrupted — saving collected data...")
|
|
|
|
learner.save_dataset(args.out)
|
|
print(f"Saved {len(learner.dataset)} samples to {args.out}")
|