Generating Data¶

For generating data from multiple farms and policies simultaneously, see Generating Data from Multiple Sources.

Generally, there are two types of data a user may want to generated: data from a prespecified policy and data from a trained RL Agent. See Prespecified Policies and RL Agent Policies for more information.

Data is generated over a range of years and range of latitudes and longitudes. These ranges can be specified by command line arguments:

python3 -m data_generation.gen_data
--lat-low 50
--lat-high 51
--lon-low 120
--lon-high 123
--year-low 1984
--year-high 2001

Additionally, two file types are supports .npz and .csv files. .npz files save the output in a (obs, action, reward, next_obs, done, info) format which is compatible for Offline RL training and loading into Replay Buffers. Meanwhile, .csv files save data in a daily observation format which can be useful for a variety of problems.

To specify a file type, include the argument:

python3 -m data_generation.gen_data
--file-type csv
OR
--file-type npz

Note

As with any script in WOFOST-Gym, the --load-data-fpath argument is allowed, meaning that data can be generated from a previously created configuration file.

Important

Latitudes are valid in the range (-90, 90)
Longitudes are valid in the range (-180, 180)
Years are valid in the range (1984, 2023)
Outside of these ranges, a missing data error will be thrown from pcse.NASAPower

Prespecified Policies¶

Many real-world agriculture policies are relatively simple and often open loop policies. That is, a valid policy may be Apply 10 kg/ha of Nitrogen fertilizer every 3 weeks.

One of the goals of the WOFOST-Gym benchmark is to support Off Policy Evaluation, and in many cases the baseline we are competing against is an open-loop policy, or a relatively simple closed-looop policy such as Apply 15 kg/ha of Nitrogen Fertilizer if the Nitrogen level in the soil drops below X amount.

To generate data from a prespecified policy, run the following:

python3 -m data_generation.gen_data --file-type npz --save-folder test --data-file test_data --policy-name Below_N

This code block will generate data in .npz format, saving it to folder test under the name test_data.npz. The generated data will be from the Below_N policy located in pcse_gym/policies.py.

Every Prespecified policy takes three specific arguments: --threshold, --interval, and --amount which control the total amount that can be applied (or the threshold below which to take the action), the interval of application, and the amount to apply of water, Nitrogen, Phosphorous, or Potassium. For more information on Prespecified Policies, see Prespecified Policies.

RL Agent Policies¶

For some use cases, we may want to use data generated from a trained RL agent to see how it performs in a variety of scenarios.

To generate data from a trained RL Agent, run the following:

python3 -m data_generation.gen_data --file-type npz --save-folder test --data-file test_data --agent-type <Agent Type> --agent_path <Path to agent.pt File>

This code block will generate data in .npz format, saving it to folder test under the name test_data.npz. The generated data will be from the specified agent.pt file.

Important

We make it very easy to specify RL agents, but this requires the RL_Agent.py file to be in a specific format. See Creating a New RL Agent for more information.
Confirm that the agent_type and agent.pt file specified are of the same agent.

Generating Data from Multiple Sources¶

There are many instances where one may want to generate data from multiple policies and farms. For instance, a farmer may want to know how her policy performs when fertilizing wheat at another farm. This is the problem of Off Policy Evaluation.

To generate data from multiple policies and farms, navigate to the gen_data_multiple.py file. For ease of use, each farm is specified by a .yaml configuration file which is created when an environment is called. See Environment Creation for more information. Modify multiple .yaml configuration files to achieve the desired site and crop configuration. Loading configurations is as follows:

config_fpaths = ["test/test1/config.yaml", "test/test2/config.yaml"]
envs = utils.make_gym_envs(args, config_fpaths=config_fpaths)

# Wrap environment with necessary wrappers
envs = [utils.wrap_env_reward(env, args) for env in envs]

Additionally, each policy and the required arguments for the policy required are specified in a list and a dictionary. Policies can either be Reinforcement Learning agents (e.g. PPO, SAC, DQN, etc) or pre-specified policies in pcse_gym.policies. For agent policies, the only argument needed is is the relative path to the agent. The pre-specified policies require parameters. See Prespecified Policies for more information.

An example configuration would look like:

# Pass the constructor for the policy, either a pcse_gym.policies or rl_utils.Agent class
pols = [PPO, policies.Interval_N, policies.Threshold_N]

# Specify the kwargs for the policies. If using a PyTorch policy from rl_algs, use
# the kwargs: `{"state_fpath": <path-to-agent>}
pols_kwargs = [{"state_fpath": "PPO_test/PPO/lnpkw-v0__rl_utils__1__1735688110/agent.pt"}, {"interval":7}, {"threshold":10, "amount":2}]

To run the script, type:

python3 -m data_generation.gen_data_multiple --save-folder test/multiple/ --data-file multiple

The output will be a .npz file with arrays of shape (num_farms, num_policies, ..) which can then be used for algorithm training or evaluation, or data visualization.