- nucon/model.py: constant input dimensions (zero variance in training
data) now get std=inf so they contribute 0 to normalised kNN distance
instead of causing catastrophic OOD from tiny float epsilon
- scripts/train_sac.py: add --load, --steps, --out CLI args; --load
hot-starts actor/critic weights from a previous run (learning_starts=0)
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- 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>
The None-param filtering probe at init also needs to wait for the game
to be reachable, not just the collection loop.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- Save dataset every N steps (default 10) so a disconnect loses at most
one checkpoint's worth of samples instead of everything
- Retry _get_state() on ConnectionError/Timeout rather than crashing,
resuming automatically once the game comes back up
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Model type is irrelevant during data collection. Models are now created
lazily on first use: train_model() creates a ReactorDynamicsModel,
fit_knn(k) creates a ReactorKNNModel. load_model() detects type by
file extension as before. drop_well_fitted() now checks model exists.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
rl.py:
- Add missing `from enum import Enum`
- Skip str-typed params in obs/action space construction (was crashing)
- Guard action space: exclude write-only (is_readable=False) and cheat params
- Fix step() param lookup (no longer iterates Nucon, uses _parameters dict directly)
- Correct sim-speed time dilation in real-game sleep
- Extract _build_param_space() helper shared by NuconEnv and NuconGoalEnv
- Add NuconGoalEnv: goal-conditioned env with normalised achieved/desired goal
vectors, compatible with SB3 HerReplayBuffer; goals sampled per episode
- Register Nucon-goal_power-v0 and Nucon-goal_temp-v0 presets
- Enum obs/action space now scalar index (not one-hot)
sim.py:
- Store self.port and self.host on NuconSimulator
- Add set_model() to accept a pre-loaded model directly
- load_model() detects type by extension (.pkl → kNN, else → NN torch)
and reads new checkpoint format with embedded input/output param lists
- _update_reactor_state() uses model.input_params (not all readable params),
calls .forward() directly for both NN and kNN, guards torch.no_grad per type
- Import ReactorKNNModel and pickle
model.py:
- save_model() embeds input_params/output_params in NN checkpoint dict
- load_model() handles new checkpoint format (state_dict key) with fallback
README.md:
- Update note: RODS_POS_ORDERED is no longer the only writable param;
game v2.2.25.213 exposes rod banks, pumps, MSCVs, switches and more
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Data collection:
- time_delta is now target game-time; wall sleep = game_delta / sim_speed
so stored deltas are uniform regardless of GAME_SIM_SPEED setting
- Auto-exclude junk params (GAME_VERSION, TIME, ALARMS_ACTIVE, …) and
params returning None (uninstalled subsystems)
- Optional include_valve_states=True adds all 53 valve positions as inputs
Model backends (model_type='nn' or 'knn'):
- ReactorKNNModel: k-nearest neighbours with GP interpolation
- Finds k nearest states, computes per-second transition rates,
linearly scales to requested game_delta (linear-in-time assumption)
- forward_with_uncertainty() returns (prediction_dict, gp_std)
where std≈0 = on known data, std≈1 = out of distribution
- NN training fixed: loss computed in tensor space, mse_loss per batch
Dataset management:
- drop_well_fitted(error_threshold): drop samples model predicts well,
keep hard cases (useful for NN curriculum)
- drop_redundant(min_state_distance, min_output_distance): drop samples
that are close in BOTH input state AND output transition space, keeping
genuinely different dynamics even at the same input state
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
