Observation spaces#

Observation space definitions#

Our goal with gym-dssat is to provide real-world like crop management problems. As a consequence, the criteria of choice to define the observations to provide to the agent are the variables to be limited in number and realistically measurables/proxied in a real crop field. The internal crop simulator –named DSSAT– has literaly hundreds of internal state variables. Based on expert knowledge, we chose a subset of DSSAT’s internal variables to be available in gym-dssat, and this subset is further restricted for the observations provided to the agent during the interactions. The Figure Successive restrictions on DSSAT state features show the succesive subsets of DSSAT’s original state features. Dashed arrows show the files defining the restrictions.

Note

For more information about how gym-dssat works, you can check out the Section Under the hood.

../../_images/stateSpace.png — Successive restrictions on DSSAT state features#

The observation space –i.e. from agent’s perspective– depend on each mode and are detailed in gym environment’s yaml configuration file. State variables can be continuous, discrete and arrays of arbitrary shapes. Actions are continuous. For instance, in the setting key of the YAML file, you can find:

#######################################################
# ACTION AND STATE SPACES RESTRICTIONS FOR ENV SETTINGS <- where settings are defined
#######################################################
 setting:
   fertilization:
     action: # action space
       - anfer
     state: # observation space
       - istage
       - vstage
       - grnwt
       - topwt
       - swfac
       - nstres
       - xlai
       - dtt
       - dap
       - cumsumfert
       - ep
     context:  # extra information the gym `info'
       - sat
       - dul
       - ll
       - dlayr
     experiment_number: 1

Hint

You can access the current values of the observation variables provided to the agent with the attribute env.observation.

Each individual variable definition, including agronomic meaning, can be found in the state key of the raw state space of the aforementioned YAML file, as exemplified below:

#################
# RAW STATE SPACE <- where all state variables imported from DSSAT-PDI are defined
#################
state:
  istage:
     type: discrete
     info: DSSAT maize growing stage (index in order 7, 8, 9, 1, 2, 3, 4, 5 ,6)
     size: 9
     # 1 = End of Juvenile Phase
     # 2 = Floral Initiation
     # 3 = Silking
     # 4 = Beginning of Grain Filling
     # 5 = End of Grain Filling
     # 6 = Maturity
     # 7 = Sowing
     # 8 = Germination
     # 9 = Emergence
  vstage:
    type: float
    low: 0
    high: 30
    info: vegetative growth stage (number of leaves)
  ...

Hint

You can access all the values of the full gym-DSSAT observation space (including variables hidden to the agent) with the attribute env._state. More information is available in the Handy attributes Section.

Getting information#

After your environment instance is initialized, you can get information about the action and observation spaces of the current mode using:

env.get_env_info()

You will get an interactive output, as exemplified below:

******************
Available actions:
******************

{'anfer': {'high': 200,
        'info': 'nitrogen to fertilize for current day (kg/ha)',
        'low': 0,
        'type': 'float'}}
press "return" to continue
{'amir': {'high': 50,
        'info': 'water depth to irrigate for current day (mm/m2)',
        'low': 0,
        'type': 'float'}}
press "return" to continue

*********************
Observation variables:
*********************

{'cleach': {'high': inf,
            'info': 'cumulative nitrate leaching (kg/ha)',
            'low': 0,
            'type': 'float'}}
press "return" to continue
...

Observation spaces

Contents

Observation spaces#

Observation space definitions#

Getting information#