greenopt-ai-gym 0.1.0

A Gymnasium environment for intelligent greenhouse control and resource optimization.

GreenOpt-AI Gym

A research-preview Python library and predefined Gymnasium environment for training and evaluating continuous-control agents against the GreenOpt-AI paper-faithful greenhouse simulator.

What is included

• Registered environment: GreenOpt/Greenhouse-v0
• Exact observation and nine-control action contracts
• Fixed training-weather scaler metadata
• Transparent multi-objective reward breakdown
• Compatibility checks for Stable-Baselines3 TD3, DDPG and SAC agents
• CLI tools and publication-ready metadata

The public Python wheel does not embed the trained .h5 files. The private complete release bundle supplied to the project owner includes the verified files under models/weights/. Confirm ownership and redistribution permissions before placing those model files in a public repository or release.

Installation

pip install greenopt-ai-gym

For the real predictive simulator:

pip install "greenopt-ai-gym[simulator]"

For Stable-Baselines3 training and validation:

pip install "greenopt-ai-gym[all]"

Required predictive model files

models/weights/
├── ghc_mlp_model_fs_model.h5
├── wcp_LSTM_model_fs_model.h5
└── rc_LSTM_model_fs_model.h5

Verified trained-model contract

The supplied trained models were inspected and tested as one complete inference chain without retraining or changing their weights:

Model	Required input	Output	Role
Climate MLP	(2016, 12)	(2016, 17)	Five-minute greenhouse-climate prediction
Crop Conv1D-LSTM	(1, 2016, 16)	(1, 3)	Weekly stem elongation, stem thickness and cumulative trusses
Resource Conv1D-LSTM	(7, 288, 10)	(7, 5)	Daily heating, high/low electricity, CO₂ and irrigation

The files were originally saved with Keras 2.8. The backend removes the obsolete time_major field from a temporary copy during modern-Keras loading. The original HDF5 files and learned weights remain unchanged.

Run the deep structural check with:

greenopt-ai validate-models models/weights

See MODEL_MANIFEST.json and FULL_MODEL_VALIDATION_REPORT.json for hashes, architectures and the end-to-end validation result.

Prediction-domain transparency

The predictive models use linear output layers. Consequently, a normalized prediction can occasionally fall outside the training interval [0, 1] even when model loading and feature ordering are correct. The simulator preserves raw predictions for paper fidelity and returns result["quality_flags"], identifying which outputs are extrapolations. A user interface should show those warnings and must not silently clip values used to calculate the original reward.

Create the environment

import gymnasium as gym
import greenopt_ai  # registers the environment

env = gym.make(
    "GreenOpt/Greenhouse-v0",
    model_dir=r"C:\path\to\models\weights",
    weather_path=r"C:\path\to\weather.csv",
    max_episode_weeks=23,
    reward_mode="paper_v1",
)

obs, info = env.reset(seed=42)
print({key: value.shape for key, value in obs.items()})
print(env.action_space.shape)  # (18144,) = 2016 five-minute rows × 9 controls

Train a TD3 agent

from stable_baselines3 import TD3
import gymnasium as gym
import greenopt_ai

env = gym.make(
    "GreenOpt/Greenhouse-v0",
    model_dir="models/weights",
    weather_path="weather.csv",
)

model = TD3(
    "MultiInputPolicy",
    env,
    learning_rate=1e-4,
    buffer_size=100,
    batch_size=256,
    gamma=0.99,
    verbose=1,
)
model.learn(total_timesteps=5000)
model.save("my_greenopt_td3_agent")

Inspect the scientific contract

greenopt-ai schema
greenopt-ai physical-ranges
greenopt-ai validate-models C:\path\to\models\weights
greenopt-ai validate-agent my_agent.zip --algorithm TD3

Observation dictionary

Key	Shape	Meaning
weather	(2016, 10)	Fixed-scaler normalized external weather for the next week
crop_params	(1, 3)	Normalized predicted crop state
resource_consumption	(1, 5)	Normalized previous weekly resource state
gh_climate	(2016, 10)	Normalized predicted greenhouse-climate state

Action space

A flat Box(0, 1, shape=(18144,)), reshaped internally to (2016, 9) in this exact control order:

1. co2_vip
2. int_white_vip
3. pH_drain_PC
4. scr_blck_vip
5. scr_enrg_vip
6. t_heat_vip
7. t_ventlee_vip
8. water_sup
9. water_sup_intervals_vip_min

Scientific boundaries

• Uploaded weather is transformed using the fixed training scaler; the package never refits a scaler per uploaded file.
• Predictive models continue to exchange normalized values internally.
• Physical-unit conversion is only for reporting and diagnostics.
• Results are simulation outcomes within the learned-data domain, not field guarantees or autonomous deployment instructions.
• GreenOpt-AI is a research simulation and decision-support environment. It is not a field-certified autonomous greenhouse controller.

See PUBLISHING_GUIDE.md before public release.