Compatibility with gymnasium / gym

The gymnasium framework in reinforcement learning is widely used. Starting from version 1.2.0 we improved the compatibility with this framework.

Starting with 1.9.1 we switch (as advised) from the legacy “gym” framework to the new “gymnasium” framework (gym is no longer maintained since v0.26.2, see https://www.gymlibrary.dev/). This change should not have any impact on older grid2op code except that you now need to use import gymnasium as gym instead of import gym in your base code.

Note

If you want to still use the “legacy” gym classes you can still do it with grid2op: Backward compatibility with openai gym is maintained.

Note

By default, if gymnasium is installed, all default classes from grid2op.gym_compat module will inherit from gymnasium. You can still retrieve the classes inheriting from gym (and not gymnasium).

More information on the section Gymnasium vs Gym

Before grid2op 1.2.0 only some classes fully implemented the open AI gym interface:

  • the grid2op.Environment (with methods such as env.reset, env.step etc.)

  • the grid2op.Agent (with the agent.act etc.)

  • the creation of pre defined environments (with grid2op.make)

Starting from 1.2.0 we implemented some automatic converters that are able to automatically map grid2op representation for the action space and the observation space into open AI gym “spaces”. More precisely these are represented as gym.spaces.Dict.

As of grid2op 1.4.0 we tighten the gap between openAI gym and grid2op by introducing the dedicated module grid2op.gym_compat . Withing this module there are lots of functionalities to convert a grid2op environment into a gym environment (that inherit gym.Env instead of “simply” implementing the open ai gym interface).

A simple usage is:

import grid2op
from grid2op.gym_compat import GymEnv

env_name = "l2rpn_case14_sandbox"  # or any other grid2op environment name
g2op_env = grid2op.make(env_name)  # create the gri2op environment

gym_env = GymEnv(g2op_env)  # create the gym environment

# check that this is a properly defined gym environment:
import gym
print(f"Is gym_env and open AI gym environment: {isinstance(gym_env, gym.Env)}")
# it shows "Is gym_env and open AI gym environment: True"

Note

To be as close as grid2op as possible, by default (using the methode discribed above) the action space will be encoded as a gymnasium.spaces.Dict with keys the attribute of a grid2op action. This might not be the best representation to perform RL with (some framework do not really like it…)

For more customization on that side, please refer to the section Customizing the action and observation space, into Box or Discrete below

Warning

The gym package has some breaking API change since its version 0.26. We attempted, in grid2op, to maintain compatibility both with former versions and later ones. This makes this class behave differently depending on the version of gym you have installed !

The main changes involve the functions env.step and env.reset (core gym functions)

This page is organized as follow:

Observation space and action space customization

By default, the action space and observation space are gym.spaces.Dict with the keys being the attribute to modify.

Default Observations space

For example, an observation space will look like:

  • “_shunt_p”: Box(env.n_shunt,) [type: float, low: -inf, high: inf]

  • “_shunt_q”: Box(env.n_shunt,) [type: float, low: -inf, high: inf]

  • “_shunt_v”: Box(env.n_shunt,) [type: float, low: -inf, high: inf]

  • “_shunt_bus”: Box(env.n_shunt,) [type: int, low: -inf, high: inf]

  • “a_ex”: Box(env.n_line,) [type: float, low: 0, high: inf]

  • “a_or”: Box(env.n_line,) [type: float, low: 0, high: inf]

  • “active_alert”: MultiBinary(env.dim_alerts)

  • “actual_dispatch”: Box(env.n_gen,)

  • “alert_duration”: Box(env.dim_alerts,) [type: int, low: 0, high: inf]

  • “attack_under_alert”: Box(env.dim_alerts,) [type: int, low: -1, high: inf]

  • “attention_budget”: Box(1,) [type: float, low: 0, high: inf]

  • “current_step”: Box(1,) [type: int, low: -inf, high: inf]

  • “curtailment”: Box(env.n_gen,) [type: float, low: 0., high: 1.0]

  • “curtailment_limit”: Box(env.n_gen,) [type: float, low: 0., high: 1.0]

  • “curtailment_limit_effective”: Box(env.n_gen,) [type: float, low: 0., high: 1.0]

  • “day”: Discrete(32)

  • “day_of_week”: Discrete(8)

  • “delta_time”: Box(0.0, inf, (1,), float32)

  • “duration_next_maintenance”: Box(env.n_line,) [type: int, low: -1, high: inf]

  • “gen_margin_down”: Box(env.n_gen,) [type: float, low: 0, high: env.gen_max_ramp_down]

  • “gen_margin_up”: Box(env.n_gen,) [type: float, low: 0, high: env.gen_max_ramp_up]

  • “gen_p”: Box(env.n_gen,) [type: float, low: env.gen_pmin, high: env.gen_pmax * 1.2]

  • “gen_p_before_curtail”: Box(env.n_gen,) [type: float, low: env.gen_pmin, high: env.gen_pmax * 1.2]

  • “gen_q”: Box(env.n_gen,) [type: float, low: -inf, high: inf]

  • “gen_theta”: Box(env.n_gen,) [type: float, low: -180, high: 180]

  • “gen_v”: Box(env.n_gen,) [type: float, low: 0, high: inf]

  • “hour_of_day”: Discrete(24)

  • “is_alarm_illegal”: Discrete(2)

  • “line_status”: MultiBinary(env.n_line)

  • “load_p”: Box(env.n_load,) [type: float, low: -inf, high: inf]

  • “load_q”: Box(env.n_load,) [type: float, low: -inf, high: inf]

  • “load_theta”: Box(env.n_load,) [type: float, low: -180, high: 180]

  • “load_v”: Box(env.n_load,) [type: float, low: -inf, high: inf]

  • “max_step”: Box(1,) [type: int, low: -inf, high: inf]

  • “minute_of_hour”: Discrete(60)

  • “month”: Discrete(13)

  • “p_ex”: Box(env.n_line,) [type: float, low: -inf, high: inf]

  • “p_or”: Box(env.n_line,) [type: float, low: -inf, high: inf]

  • “q_ex”: Box(env.n_line,) [type: float, low: -inf, high: inf]

  • “q_or”: Box(env.n_line,) [type: float, low: -inf, high: inf]

  • “rho”: Box(env.n_line,) [type: float, low: 0., high: inf]

  • “storage_charge”: Box(env.n_storage,) [type: float, low: 0., high: env.storage_Emax]

  • “storage_power”: Box(env.n_storage,) [type: float, low: -env.storage_max_p_prod, high: env.storage_max_p_absorb]

  • “storage_power_target”: Box(env.n_storage,) [type: float, low: -env.storage_max_p_prod, high: env.storage_max_p_absorb]

  • “storage_theta”: Box(env.n_storage,) [type: float, low: -180., high: 180.]

  • “target_dispatch”: Box(env.n_gen,) [type: float, low: -inf, high: inf]

  • “thermal_limit”: Box(env.n_line,) [type: int, low: 0, high: inf]

  • “theta_ex”: Box(env.n_line,) [type: float, low: -180., high: 180.]

  • “theta_or”: Box(env.n_line,) [type: float, low: -180., high: 180.]

  • “time_before_cooldown_line”: Box(env.n_line,) [type: int, low: 0, high: depending on parameters]

  • “time_before_cooldown_sub”: Box(env.n_sub,) [type: int, low: 0, high: depending on parameters]

  • “time_next_maintenance”: Box(env.n_line,) [type: int, low: 0, high: inf]

  • “time_since_last_alarm”: Box(1,) [type: int, low: -1, high: inf]

  • “time_since_last_alert”: Box(env.dim_alerts,) [type: int, low: -1, high: inf]

  • “time_since_last_attack”: Box(env.dim_alerts,) [type: int, low: -1, high: inf]

  • “timestep_overflow”: Box(env.n_line,) [type: int, low: 0, high: inf]

  • “topo_vect”: Box(env.dim_topo,) [type: int, low: -1, high: 2]

  • “total_number_of_alert”: Box(1 if env.dim_alerts > 0 else 0,) [type: int, low: 0, high: inf]

  • “v_ex”: Box(env.n_line,) [type: float, low: 0, high: inf]

  • “v_or”: Box(env.n_line,) [type: flaot, low: 0, high: inf]

  • “was_alarm_used_after_game_over”: Discrete(2)

  • “was_alert_used_after_attack”: Box(env.dim_alerts,) [type: int, low: -1, high: 1]

  • “year”: Discrete(2100)

Each keys correspond to an attribute of the observation. In this example “line_status”: MultiBinary(20) represents the attribute obs.line_status which is a boolean vector (for each powerline True encodes for “connected” and False for “disconnected”) See the chapter Observation for more information about these attributes.

You can transform the observation space as you wish. There are some examples in the notebooks.

Default Action space

The default action space is also a type of gym Dict. As for the observation space above, it is a straight translation from the attribute of the action to the key of the dictionary. This gives:

  • “change_bus”: MultiBinary(env.dim_topo)

  • “change_line_status”: MultiBinary(env.n_line)

  • “curtail”: Box(env.n_gen) [type: float, low=0., high=1.0]

  • “redispatch”: Box(env.n_gen) [type: float, low=-env.gen_max_ramp_down, high=`env.gen_max_ramp_up`]

  • “set_bus”: Box(env.dim_topo) [type: int, low=-1, high=2]

  • “set_line_status”: Box(env.n_line) [type: int, low=-1, high=1]

  • “storage_power”: Box(env.n_storage) [type: float, low=-env.storage_max_p_prod, high=`env.storage_max_p_absorb`]

  • “raise_alarm”: MultiBinary(env.dim_alarms)

  • “raise_alert”: MultiBinary(env.dim_alerts)

For example you can create a “gym action” (for the default encoding) like:

import grid2op
from grid2op.gym_compat import GymEnv
import numpy as np

env_name = ...

env = grid2op.make(env_name)
gym_env = GymEnv(env)

seed = ...
obs, info = gym_env.reset(seed)  # for new gym interface

# do nothing
gym_act = {}
obs, reward, done, truncated, info = gym_env.step(gym_act)

#change the bus of the element 6 and 7 of the "topo_vect"
gym_act = {}
gym_act["change_bus"] = np.zeros(env.dim_topo, dtype=np.int8)   # gym encoding of a multi binary
gym_act["change_bus"][[6, 7]] = 1
obs, reward, done, truncated, info = gym_env.step(gym_act)

# redispatch generator 2 of 1.7MW
gym_act = {}
gym_act["redispatch"] = np.zeros(env.n_gen, dtype=np.float32)   # gym encoding of a Box
gym_act["redispatch"][2] = 1.7
obs, reward, done, truncated, info = gym_env.step(gym_act)

# set the bus of element 8 and 9 to bus 2
gym_act = {}
gym_act["set_bus"] = np.zeros(env.dim_topo, dtype=int)   # gym encoding of a Box
gym_act["set_bus"][[8, 9]] = 2
obs, reward, done, truncated, info = gym_env.step(gym_act)

# of course, you can set_bus, redispatch, change the storage units etc. in the same action.

This way of doing things is perfectly grounded. It works but it is quite verbose and not really “ML friendly”. You can customize the way you “encode” your actions / observations relatively easily. Some examples are given in the following subsections.

Customizing the action and observation space

We offer some convenience functions to customize these spaces.

If you want a full control on this spaces, you need to implement something like:

import grid2op
env_name = ...
env = grid2op.make(env_name)

from grid2op.gym_compat import GymEnv
# this of course will not work... Replace "AGymSpace" with a normal gym space, like Dict, Box, MultiDiscrete etc.
from gym.spaces import AGymSpace
gym_env = GymEnv(env)

class MyCustomObservationSpace(AGymSpace):
    def __init__(self, whatever, you, want):
        # do as you please here
        pass
        # don't forget to initialize the base class
        AGymSpace.__init__(self, see, gym, doc, as, to, how, to, initialize, it)
        # eg. Box.__init__(self, low=..., high=..., dtype=float)

    def to_gym(self, observation):
        # this is this very same function that you need to implement
        # it should have this exact name, take only one observation (grid2op) as input
        # and return a gym object that belong to your space "AGymSpace"
        return SomethingThatBelongTo_AGymSpace
        # eg. return np.concatenate((obs.gen_p * 0.1, np.sqrt(obs.load_p))

gym_env.observation_space = MyCustomObservationSpace(whatever, you, wanted)

And for the action space:

import grid2op
env_name = ...
env = grid2op.make(env_name)

from grid2op.gym_compat import GymEnv
# this of course will not work... Replace "AGymSpace" with a normal gym space, like Dict, Box, MultiDiscrete etc.
from gym.spaces import AGymSpace
gym_env = GymEnv(env)

class MyCustomActionSpace(AGymSpace):
    def __init__(self, whatever, you, want):
        # do as you please here
        pass
        # don't forget to initialize the base class
        AGymSpace.__init__(self, see, gym, doc, as, to, how, to, initialize, it)
        # eg. MultiDiscrete.__init__(self, nvec=...)

    def from_gym(self, gym_action):
        # this is this very same function that you need to implement
        # it should have this exact name, take only one action (member of your gym space) as input
        # and return a grid2op action
        return TheGymAction_ConvertedTo_Grid2op_Action
        # eg. return np.concatenate((obs.gen_p * 0.1, np.sqrt(obs.load_p))

gym_env.action_space = MyCustomActionSpace(whatever, you, wanted)

There are some pre defined transformation (for example transforming the action to Discrete or MultiDiscrete). Do not hesitate to have a look at the section Customizing the action and observation space, into Box or Discrete.

Some already implemented customization

However, if you don’t want to fully customize everything, we encourage you to have a look at the “GymConverter” that we coded to ease this process.

They all more or less the same manner. We show here an example of a “converter” that will scale the data (removing the value in substract and divide input data by divide):

import grid2op
from grid2op.gym_compat import GymEnv
from grid2op.gym_compat import ScalerAttrConverter

env_name = "l2rpn_case14_sandbox"  # or any other grid2op environment name
g2op_env = grid2op.make(env_name)  # create the gri2op environment

gym_env = GymEnv(g2op_env)  # create the gym environment

ob_space = gym_env.observation_space
ob_space = ob_space.reencode_space("actual_dispatch",
                                   ScalerAttrConverter(substract=0.,
                                                       divide=env.gen_pmax,
                                                       init_space=ob_space["actual_dispatch"]
                                                       )
                                   )

gym_env.observation_space = ob_space

You can also add a specific keys into this observation space, for example say you want to compute the log of the loads instead of giving the direct value to your agent. This can be done with:

import grid2op
from grid2op.gym_compat import GymEnv
from grid2op.gym_compat import ScalerAttrConverter

env_name = "l2rpn_case14_sandbox"  # or any other grid2op environment name
g2op_env = grid2op.make(env_name)  # create the gri2op environment

gym_env = GymEnv(g2op_env)  # create the gym environment

ob_space = gym_env.observation_space
shape_ = (g2op_env.n_load, )
ob_space = ob_space.add_key("log_load",
                             lambda obs: np.log(obs.load_p),
                                      Box(shape=shape_,
                                          low=np.full(shape_, fill_value=-np.inf, dtype=float),
                                          high=np.full(shape_, fill_value=-np.inf, dtype=float),
                                          dtype=float
                                          )
                                   )

gym_env.observation_space = ob_space
# and now you will get the key "log_load" as part of your gym observation.

A detailed list of such “converter” is documented on the section “Detailed Documentation by class”. In the table below we describe some of them (nb if you notice a converter is not displayed there, do not hesitate to write us a “feature request” for the documentation, thanks in advance)

Converter name

Objective

ContinuousToDiscreteConverter

Convert a continuous space into a discrete one

MultiToTupleConverter

Convert a gym MultiBinary to a gym Tuple of gym Binary and a gym MultiDiscrete to a Tuple of Discrete

ScalerAttrConverter

Allows to scale (divide an attribute by something and subtract something from it)

BaseGymSpaceConverter.add_key

Allows you to compute another “part” of the observation space (you add an information to the gym space)

BaseGymSpaceConverter.keep_only_attr

Allows you to specify which part of the action / observation you want to keep

BaseGymSpaceConverter.ignore_attr

Allows you to ignore some attributes of the action / observation (they will not be part of the gym space)

Warning

TODO: Help more than welcome !

Organize this page with a section for each “use”:

  • scale de the data

  • keep only some part of the observation

  • add some info to the observation

  • transform a box to a discrete action space

  • use MultiDiscrete

Instead of having the current ordering of things

Note

With the “converters” above, note that the observation space AND action space will still inherit from gym Dict.

They are complex spaces that are not well handled by some RL framework.

These converters only change the keys of these dictionaries !

Customizing the action and observation space, into Box or Discrete

The use of the converter above is nice if you can work with gym Dict, but in some cases, or for some frameworks it is not convenient to do it at all.

TO alleviate this problem, we developed 4 types of gym action space, following the architecture detailed in subsection Customizing the action and observation space

Converter name

Objective

BoxGymObsSpace

Convert the observation space to a single “Box”

BoxGymActSpace

Convert a gym MultiBinary to a gym Tuple of gym Binary and a gym MultiDiscrete to a Tuple of Discrete

MultiDiscreteActSpace

Allows to scale (divide an attribute by something and subtract something from it)

DiscreteActSpace

Allows you to compute another “part” of the observation space (you add an information to the gym space)

They can all be used like:

import grid2op
env_name = ...
env = grid2op.make(env_name)

from grid2op.gym_compat import GymEnv, BoxGymObsSpace
gym_env = GymEnv(env)
gym_env.observation_space = BoxGymObsSpace(gym_env.init_env)
gym_env.action_space = MultiDiscreteActSpace(gym_env.init_env)

We encourage you to visit the documentation for more information on how to use these classes. Each offer different possible customization.

Gymnasium vs Gym

Starting from grid2op 1.9.1 we introduced the compatibility with gymnasium package (the replacement of the gym package that will no longer be maintained).

By default, if gymnasium is installed on your machine, all classes from the grid2op.gym_compat module will inherit from gymnasium. That is GymEnv will be inherit from gymnasium.Env`(and not `gym.Env), GymActionSpace will inherit from gymnasium.spaces.Dict (and not from gym.spaces.Dict) etc.

But we wanted to maintain Backward compatibility. It is ensured in two different ways:

  1. if you have both gymnasium and gym installed on your machine, you can choose which “framework” you want to use by explicitly using the right grid2op class. For example, if you want a gym environment (inheriting from gym.Env) you can use GymEnv_Modern`and if you want to explicitly stay in `gymnasium you can use GymnasiumEnv

  2. if you don’t want to have gymnasium and only gym is installed then the default grid2op class will stay in the gym eco system. In this case, gym.Env will be GymEnv_Modern and all the code previously written will work exactly as before.

Note

As you understood if you want to keep the behaviour of grid2op prior to 1.9.1 the simplest solution would be not to install gymnasium at all.

If however you want to benefit from the latest gymnasium package, you can keep the previous code you have and simply install gymnasium. All classes defined there will still be defined and you will be able to use gymnasium transparently.

The table bellow summarize the correspondance between the default classes and the classes specific to gymnasium / gym:

Default class

Class with gymnasium

Class with gym

BaseGymAttrConverter

BaseGymnasiumAttrConverter

BaseLegacyGymAttrConverter

BoxGymActSpace

BoxGymnasiumActSpace

BoxLegacyGymActSpace

BoxGymObsSpace

BoxGymnasiumObsSpace

BoxLegacyGymObsSpace

ContinuousToDiscreteConverter

ContinuousToDiscreteConverterGymnasium

ContinuousToDiscreteConverterLegacyGym

DiscreteActSpace

DiscreteActSpaceGymnasium

DiscreteActSpaceLegacyGym

GymActionSpace

GymnasiumActionSpace

LegacyGymActionSpace

GymObservationSpace

GymnasiumObservationSpace

LegacyGymObservationSpace

_BaseGymSpaceConverter

_BaseGymnasiumSpaceConverter

_BaseLegacyGymSpaceConverter

GymEnv

GymnasiumEnv

GymEnv_Modern / GymEnv_Legacy

MultiToTupleConverter

MultiToTupleConverterGymnasium

MultiToTupleConverterLegacyGym

MultiDiscreteActSpace

MultiDiscreteActSpaceGymnasium

MultiDiscreteActSpaceLegacyGym

ScalerAttrConverter

ScalerAttrConverterGymnasium

ScalerAttrConverterLegacyGym

Troubleshoot with some frameworks

Python complains about pickle

This usually takes the form of an error with XXX_env_name (eg CompleteObservation_l2rpn_wcci_2022) is not serializable.

This is because grid2op will (to save computation time) generate some classes (the classes themseleves) on the fly, once the environment is loaded. And unfortunately, pickle module is not always able to process these (meta) data.

Try to first create (automatically!) the files containing the description of the classes used by your environment (for example):

from grid2op import make
from grid2op.Reward import RedispReward
from lightsim2grid import LightSimBackend

env_name = 'l2rpn_wcci_2022'
backend_class = LightSimBackend
env = make(env_name, reward_class=RedispReward, backend=backend_class())
env.generate_classes()

Note

This piece of code is to do once (each time you change the backend or the env name)

And then proceed as usual by loading the grid2op environment with the key-word experimental_read_from_local_dir

from grid2op import make
from grid2op.Reward import RedispReward
from lightsim2grid import LightSimBackend

env_name = 'l2rpn_wcci_2022'
backend_class = LightSimBackend
env = make(env_name, reward_class=RedispReward, backend=backend_class(),
           experimental_read_from_local_dir=True)
# do whatever

Observation XXX outside given space YYY

Often encountered with ray[rllib] this is due to a technical aspect (slack bus) of the grid which may cause issue with gen_p being above / bellow pmin / pmax for certain generators.

You can get rid of it by modifying the observation space and “remove” the low / high values on pmin and pmax:

# we suppose you already have an observation space
self.observation_space["gen_p"].low[:] = -np.inf
self.observation_space["gen_p"].high[:] = np.inf

Detailed Documentation by class

Classes:

BaseGymAttrConverter

alias of BaseGymnasiumAttrConverter

BaseGymnasiumAttrConverter([space, ...])

TODO work in progress !

BaseLegacyGymAttrConverter([space, ...])

TODO work in progress !

BoxGymActSpace

alias of BoxGymnasiumActSpace

BoxGymObsSpace

alias of BoxGymnasiumObsSpace

BoxGymnasiumActSpace(grid2op_action_space[, ...])

This class allows to convert a grid2op action space into a gym "Box" which is a regular Box in R^d.

BoxGymnasiumObsSpace(grid2op_observation_space)

This class allows to convert a grid2op observation space into a gym "Box" which is a regular Box in R^d.

BoxLegacyGymActSpace(grid2op_action_space[, ...])

This class allows to convert a grid2op action space into a gym "Box" which is a regular Box in R^d.

BoxLegacyGymObsSpace(grid2op_observation_space)

This class allows to convert a grid2op observation space into a gym "Box" which is a regular Box in R^d.

ContinuousToDiscreteConverter

alias of ContinuousToDiscreteConverterGymnasium

ContinuousToDiscreteConverterGymnasium(nb_bins)

TODO doc in progress

ContinuousToDiscreteConverterLegacyGym(nb_bins)

TODO doc in progress

DiscreteActSpace

alias of DiscreteActSpaceGymnasium

DiscreteActSpaceGymnasium(grid2op_action_space)

TODO the documentation of this class is in progress.

DiscreteActSpaceLegacyGym(grid2op_action_space)

TODO the documentation of this class is in progress.

GymActionSpace

alias of GymnasiumActionSpace

GymEnv

alias of GymnasiumEnv

GymEnv_Legacy(env_init[, shuffle_chronics, ...])

fully implements the openAI gym API by using the GymActionSpace and GymObservationSpace for compliance with openAI gym.

GymEnv_Modern(env_init[, shuffle_chronics, ...])

fully implements the openAI gym API by using the GymActionSpace and GymObservationSpace for compliance with openAI gym.

GymObservationSpace

alias of GymnasiumObservationSpace

GymnasiumActionSpace(env[, converter, ...])

This class enables the conversion of the action space into a gym "space".

GymnasiumEnv(env_init[, shuffle_chronics, ...])

fully implements the openAI gym API by using the GymActionSpace and GymObservationSpace for compliance with openAI gym.

GymnasiumObservationSpace(env[, dict_variables])

TODO explain gym / gymnasium

LegacyGymActionSpace(env[, converter, ...])

This class enables the conversion of the action space into a gym "space".

LegacyGymObservationSpace(env[, dict_variables])

TODO explain gym / gymnasium

MultiDiscreteActSpace

alias of MultiDiscreteActSpaceGymnasium

MultiDiscreteActSpaceGymnasium(...[, ...])

This class allows to convert a grid2op action space into a gym "MultiDiscrete".

MultiDiscreteActSpaceLegacyGym(...[, ...])

This class allows to convert a grid2op action space into a gym "MultiDiscrete".

MultiToTupleConverter

alias of MultiToTupleConverterGymnasium

MultiToTupleConverterGymnasium([init_space])

Some framework, for example ray[rllib] do not support MultiBinary nor MultiDiscrete gym action space.

MultiToTupleConverterLegacyGym([init_space])

Some framework, for example ray[rllib] do not support MultiBinary nor MultiDiscrete gym action space.

ScalerAttrConverter

alias of ScalerAttrConverterGymnasium

ScalerAttrConverterGymnasium(substract, divide)

This is a scaler that transforms a initial gym space init_space into its scale version.

ScalerAttrConverterLegacyGym(substract, divide)

This is a scaler that transforms a initial gym space init_space into its scale version.

grid2op.gym_compat.BaseGymAttrConverter

alias of BaseGymnasiumAttrConverter

class grid2op.gym_compat.BaseGymnasiumAttrConverter(space=None, gym_to_g2op=None, g2op_to_gym=None)

TODO work in progress !

Need help if you can :-)

Warning

Depending on the presence absence of gymnasium and gym packages this class might behave differently.

In grid2op we tried to maintain compatibility both with gymnasium (newest) and gym (legacy, no more maintained) RL packages. The behaviour is the following:

See Gymnasium vs Gym for more information

class grid2op.gym_compat.BaseLegacyGymAttrConverter(space=None, gym_to_g2op=None, g2op_to_gym=None)

TODO work in progress !

Need help if you can :-)

Warning

Depending on the presence absence of gymnasium and gym packages this class might behave differently.

In grid2op we tried to maintain compatibility both with gymnasium (newest) and gym (legacy, no more maintained) RL packages. The behaviour is the following:

See Gymnasium vs Gym for more information

grid2op.gym_compat.BoxGymActSpace

alias of BoxGymnasiumActSpace

grid2op.gym_compat.BoxGymObsSpace

alias of BoxGymnasiumObsSpace

class grid2op.gym_compat.BoxGymnasiumActSpace(grid2op_action_space: ActionSpace, attr_to_keep: Tuple[Literal['set_line_status'], Literal['change_line_status'], Literal['set_bus'], Literal['change_bus'], Literal['redispatch'], Literal['set_storage'], Literal['curtail'], Literal['curtail_mw'], Literal['raise_alarm'], Literal['raise_alert']] | None = ('redispatch', 'set_storage', 'curtail'), add: Dict[str, Any] | None = None, multiply: Dict[str, Any] | None = None, functs: Dict[str, Any] | None = None)

This class allows to convert a grid2op action space into a gym “Box” which is a regular Box in R^d.

It also allows to customize which part of the action you want to use and offer capacity to center / reduce the data or to use more complex function from the observation.

Note

Though it is possible to use every type of action with this type of action space, be aware that this is not recommended at all to use it for discrete attribute (set_bus, change_bus, set_line_status or change_line_status) !

Basically, when doing action in gym for these attributes, this converter will involve rounding and is definitely not the best representation. Prefer the MultiDiscreteActSpace or the DiscreteActSpace classes.

Note

A gymnasium Box is encoded as a numpy array, see the example section for more information.

Danger

If you use this encoding for the “curtailment” you might end up with “weird” behaviour. Your agent will perfom some kind of curtailment at each step (there is no real way to express “I don’t want to curtail”) So the index corresponding to the “curtail” type of actions should rather be “1.” (meaning “limit the value at 100%” which is somewhat equivalent to “I don’t want to curtail”)

Examples

If you simply want to use it you can do:

import grid2op
env_name = "l2rpn_case14_sandbox"  # or any other name
env = grid2op.make(env_name)

from grid2op.gym_compat import GymEnv, BoxGymActSpace
gym_env = GymEnv(env)

gym_env.action_space = BoxGymActSpace(env.action_space)

In this case it will extract all the features in all the action (a detailed list is given in the documentation at Action).

You can select the attribute you want to keep, for example:

gym_env.action_space = BoxGymActSpace(env.action_space,
                                           attr_to_keep=['redispatch', "curtail"])

You can also apply some basic transformation to the attribute of the action. This can be done with:

gym_env.action_space = BoxGymActSpace(env.action_space,
                                      attr_to_keep=['redispatch', "curtail"],
                                      multiply={"redispatch": env.gen_max_ramp_up},
                                      add={"redispatch": 0.5 * env.gen_max_ramp_up})

In the above example, the resulting “redispatch” part of the vector will be given by the following formula: grid2op_act = gym_act * multiply + add

Hint: you can use: multiply being the standard deviation and add being the average of the attribute.

Warning

Depending on the presence absence of gymnasium and gym packages this class might behave differently.

In grid2op we tried to maintain compatibility both with gymnasium (newest) and gym (legacy, no more maintained) RL packages. The behaviour is the following:

See Gymnasium vs Gym for more information

For the “l2rpn_case14_sandbox” environment, a code using BoxGymActSpace can look something like (if you want to build action “by hands”):

import grid2op
from grid2op.gym_compat import GymEnv, BoxGymActSpace
import numpy as np
env_name = "l2rpn_case14_sandbox"

env = grid2op.make(env_name)
gym_env =  GymEnv(env)
gym_env.action_space = BoxGymActSpace(env.action_space)

obs = gym_env.reset()  # obs will be an OrderedDict (default, but you can customize it)

# you can do a "do nothing" action
act = np.zeros(gym_env.action_space.shape)
# see danger about curtailment !
start_, end_ = gym_env.action_space.get_indexes("curtail")
## real version, not in the space... (write an issue if it's a problem for you)
act[start_:end_] = -1
## version "in the space"
# act[start_:end_] = 1
print(gym_env.action_space.from_gym(act))
obs, reward, done, truncated, info = gym_env.step(act)

# you can also do a random action:
act = gym_env.action_space.sample()
print(gym_env.action_space.from_gym(act))
obs, reward, done, truncated, info = gym_env.step(act)

# you can do an action on say redispatch (for example)
act = np.zeros(gym_env.action_space.shape)
key = "redispatch"  # "redispatch", "curtail", "set_storage" (but there is no storage on this env)
start_, end_ = gym_env.action_space.get_indexes(key)
act[start_:end_] = np.random.uniform(high=1, low=-1, size=env.gen_redispatchable.sum())  # the dispatch vector
print(gym_env.action_space.from_gym(act))
obs, reward, done, truncated, info = gym_env.step(act)

Notes

For more customization, this code is roughly equivalent to something like:

import grid2op
env_name = "l2rpn_case14_sandbox"  # or any other name
env = grid2op.make(env_name)

from grid2op.gym_compat import GymEnv
# this of course will not work... Replace "AGymSpace" with a normal gym space, like Dict, Box, MultiDiscrete etc.
from gym.spaces import AGymSpace
gym_env = GymEnv(env)

class MyCustomActionSpace(AGymSpace):
    def __init__(self, whatever, you, want):
        # do as you please here
        pass
        # don't forget to initialize the base class
        AGymSpace.__init__(self, see, gym, doc, as, to, how, to, initialize, it)
        # eg. MultiDiscrete.__init__(self, nvec=...)

    def from_gym(self, gym_action):
        # this is this very same function that you need to implement
        # it should have this exact name, take only one action (member of your gym space) as input
        # and return a grid2op action
        return TheGymAction_ConvertedTo_Grid2op_Action
        # eg. return np.concatenate((obs.gen_p * 0.1, np.sqrt(obs.load_p))

gym_env.action_space.close()
gym_env.action_space = MyCustomActionSpace(whatever, you, wanted)

And you can implement pretty much anything in the “from_gym” function.

class grid2op.gym_compat.BoxGymnasiumObsSpace(grid2op_observation_space, attr_to_keep=('year', 'month', 'day', 'hour_of_day', 'minute_of_hour', 'day_of_week', 'gen_p', 'gen_p_before_curtail', 'gen_q', 'gen_v', 'gen_margin_up', 'gen_margin_down', 'load_p', 'load_q', 'load_v', 'p_or', 'q_or', 'v_or', 'a_or', 'p_ex', 'q_ex', 'v_ex', 'a_ex', 'rho', 'line_status', 'timestep_overflow', 'topo_vect', 'time_before_cooldown_line', 'time_before_cooldown_sub', 'time_next_maintenance', 'duration_next_maintenance', 'target_dispatch', 'actual_dispatch', 'storage_charge', 'storage_power_target', 'storage_power', 'curtailment', 'curtailment_limit', 'curtailment_limit_effective', 'thermal_limit', 'is_alarm_illegal', 'time_since_last_alarm', 'last_alarm', 'attention_budget', 'was_alarm_used_after_game_over', 'max_step', 'active_alert', 'attack_under_alert', 'time_since_last_alert', 'alert_duration', 'total_number_of_alert', 'time_since_last_attack', 'was_alert_used_after_attack', 'theta_or', 'theta_ex', 'load_theta', 'gen_theta'), subtract=None, divide=None, functs=None)

This class allows to convert a grid2op observation space into a gym “Box” which is a regular Box in R^d.

It also allows to customize which part of the observation you want to use and offer capacity to center / reduce the data or to use more complex function from the observation.

Warning

Depending on the presence absence of gymnasium and gym packages this class might behave differently.

In grid2op we tried to maintain compatibility both with gymnasium (newest) and gym (legacy, no more maintained) RL packages. The behaviour is the following:

See Gymnasium vs Gym for more information

Note

A gymnasium Box is encoded as a numpy array.

Examples

If you simply want to use it you can do:

import grid2op
env_name = "l2rpn_case14_sandbox"  # or any other name
env = grid2op.make(env_name)

from grid2op.gym_compat import GymEnv, BoxGymObsSpace
gym_env = GymEnv(env)

gym_env.observation_space = BoxGymObsSpace(env.observation_space)

In this case it will extract all the features in all the observation (a detailed list is given in the documentation at Observation.

You can select the attribute you want to keep, for example:

gym_env.observation_space = BoxGymObsSpace(env.observation_space,
                                           attr_to_keep=['load_p', "gen_p", "rho])

You can also apply some basic transformation to the attribute of the observation before building the resulting gym observation (which in this case is a vector). This can be done with:

gym_env.observation_space = BoxGymObsSpace(env.observation_space,
                                           attr_to_keep=['load_p', "gen_p", "rho"],
                                           divide={"gen_p": env.gen_pmax},
                                           substract={"gen_p": 0.5 * env.gen_pmax})

In the above example, the resulting “gen_p” part of the vector will be given by the following formula: gym_obs = (grid2op_obs - substract) / divide.

Hint: you can use: divide being the standard deviation and subtract being the average of the attribute on a few episodes for example. This can be done with grid2op.utils.EpisodeStatistics for example.

Finally, you can also modify more the attribute of the observation and add it to your box. This can be done rather easily with the “functs” argument like:

gym_env.observation_space = BoxGymObsSpace(env.observation_space,
                                           attr_to_keep=["connectivity_matrix", "log_load"],
                                           functs={"connectivity_matrix":
                                                      (lambda grid2opobs: grid2opobs.connectivity_matrix().flatten(),
                                                       0., 1.0, None, None),
                                                   "log_load":
                                                    (lambda grid2opobs: np.log(grid2opobs.load_p),
                                                    None, 10., None, None)
                                                }
                                           )

In this case, “functs” should be a dictionary, the “keys” should be string (keys should also be present in the attr_to_keep list) and the values should count 5 elements (callable, low, high, shape, dtype) with:

  • callable a function taking as input a grid2op observation and returning a numpy array

  • low (optional) (put None if you don’t want to specify it, defaults to -np.inf) the lowest value your numpy array can take. It can be a single number or an array with the same shape as the return value of your function.

  • high (optional) (put None if you don’t want to specify it, defaults to np.inf) the highest value your numpy array can take. It can be a single number or an array with the same shape as the return value of your function.

  • shape (optional) (put None if you don’t want to specify it) the shape of the return value of your function. It should be a tuple (and not a single number). By default it is computed with by applying your function to an observation.

  • dtype (optional, put None if you don’t want to change it, defaults to np.float32) the type of the numpy array as output of your function.

Notes

The range of the values for “gen_p” / “prod_p” are not strictly env.gen_pmin and env.gen_pmax. This is due to the “approximation” when some redispatching is performed (the precision of the algorithm that computes the actual dispatch from the information it receives) and also because sometimes the losses of the grid are really different that the one anticipated in the “chronics” (yes env.gen_pmin and env.gen_pmax are not always ensured in grid2op)

class grid2op.gym_compat.BoxLegacyGymActSpace(grid2op_action_space: ActionSpace, attr_to_keep: Tuple[Literal['set_line_status'], Literal['change_line_status'], Literal['set_bus'], Literal['change_bus'], Literal['redispatch'], Literal['set_storage'], Literal['curtail'], Literal['curtail_mw'], Literal['raise_alarm'], Literal['raise_alert']] | None = ('redispatch', 'set_storage', 'curtail'), add: Dict[str, Any] | None = None, multiply: Dict[str, Any] | None = None, functs: Dict[str, Any] | None = None)

This class allows to convert a grid2op action space into a gym “Box” which is a regular Box in R^d.

It also allows to customize which part of the action you want to use and offer capacity to center / reduce the data or to use more complex function from the observation.

Note

Though it is possible to use every type of action with this type of action space, be aware that this is not recommended at all to use it for discrete attribute (set_bus, change_bus, set_line_status or change_line_status) !

Basically, when doing action in gym for these attributes, this converter will involve rounding and is definitely not the best representation. Prefer the MultiDiscreteActSpace or the DiscreteActSpace classes.

Note

A gymnasium Box is encoded as a numpy array, see the example section for more information.

Danger

If you use this encoding for the “curtailment” you might end up with “weird” behaviour. Your agent will perfom some kind of curtailment at each step (there is no real way to express “I don’t want to curtail”) So the index corresponding to the “curtail” type of actions should rather be “1.” (meaning “limit the value at 100%” which is somewhat equivalent to “I don’t want to curtail”)

Examples

If you simply want to use it you can do:

import grid2op
env_name = "l2rpn_case14_sandbox"  # or any other name
env = grid2op.make(env_name)

from grid2op.gym_compat import GymEnv, BoxGymActSpace
gym_env = GymEnv(env)

gym_env.action_space = BoxGymActSpace(env.action_space)

In this case it will extract all the features in all the action (a detailed list is given in the documentation at Action).

You can select the attribute you want to keep, for example:

gym_env.action_space = BoxGymActSpace(env.action_space,
                                           attr_to_keep=['redispatch', "curtail"])

You can also apply some basic transformation to the attribute of the action. This can be done with:

gym_env.action_space = BoxGymActSpace(env.action_space,
                                      attr_to_keep=['redispatch', "curtail"],
                                      multiply={"redispatch": env.gen_max_ramp_up},
                                      add={"redispatch": 0.5 * env.gen_max_ramp_up})

In the above example, the resulting “redispatch” part of the vector will be given by the following formula: grid2op_act = gym_act * multiply + add

Hint: you can use: multiply being the standard deviation and add being the average of the attribute.

Warning

Depending on the presence absence of gymnasium and gym packages this class might behave differently.

In grid2op we tried to maintain compatibility both with gymnasium (newest) and gym (legacy, no more maintained) RL packages. The behaviour is the following:

See Gymnasium vs Gym for more information

For the “l2rpn_case14_sandbox” environment, a code using BoxGymActSpace can look something like (if you want to build action “by hands”):

import grid2op
from grid2op.gym_compat import GymEnv, BoxGymActSpace
import numpy as np
env_name = "l2rpn_case14_sandbox"

env = grid2op.make(env_name)
gym_env =  GymEnv(env)
gym_env.action_space = BoxGymActSpace(env.action_space)

obs = gym_env.reset()  # obs will be an OrderedDict (default, but you can customize it)

# you can do a "do nothing" action
act = np.zeros(gym_env.action_space.shape)
# see danger about curtailment !
start_, end_ = gym_env.action_space.get_indexes("curtail")
## real version, not in the space... (write an issue if it's a problem for you)
act[start_:end_] = -1
## version "in the space"
# act[start_:end_] = 1
print(gym_env.action_space.from_gym(act))
obs, reward, done, truncated, info = gym_env.step(act)

# you can also do a random action:
act = gym_env.action_space.sample()
print(gym_env.action_space.from_gym(act))
obs, reward, done, truncated, info = gym_env.step(act)

# you can do an action on say redispatch (for example)
act = np.zeros(gym_env.action_space.shape)
key = "redispatch"  # "redispatch", "curtail", "set_storage" (but there is no storage on this env)
start_, end_ = gym_env.action_space.get_indexes(key)
act[start_:end_] = np.random.uniform(high=1, low=-1, size=env.gen_redispatchable.sum())  # the dispatch vector
print(gym_env.action_space.from_gym(act))
obs, reward, done, truncated, info = gym_env.step(act)

Notes

For more customization, this code is roughly equivalent to something like:

import grid2op
env_name = "l2rpn_case14_sandbox"  # or any other name
env = grid2op.make(env_name)

from grid2op.gym_compat import GymEnv
# this of course will not work... Replace "AGymSpace" with a normal gym space, like Dict, Box, MultiDiscrete etc.
from gym.spaces import AGymSpace
gym_env = GymEnv(env)

class MyCustomActionSpace(AGymSpace):
    def __init__(self, whatever, you, want):
        # do as you please here
        pass
        # don't forget to initialize the base class
        AGymSpace.__init__(self, see, gym, doc, as, to, how, to, initialize, it)
        # eg. MultiDiscrete.__init__(self, nvec=...)

    def from_gym(self, gym_action):
        # this is this very same function that you need to implement
        # it should have this exact name, take only one action (member of your gym space) as input
        # and return a grid2op action
        return TheGymAction_ConvertedTo_Grid2op_Action
        # eg. return np.concatenate((obs.gen_p * 0.1, np.sqrt(obs.load_p))

gym_env.action_space.close()
gym_env.action_space = MyCustomActionSpace(whatever, you, wanted)

And you can implement pretty much anything in the “from_gym” function.

class grid2op.gym_compat.BoxLegacyGymObsSpace(grid2op_observation_space, attr_to_keep=('year', 'month', 'day', 'hour_of_day', 'minute_of_hour', 'day_of_week', 'gen_p', 'gen_p_before_curtail', 'gen_q', 'gen_v', 'gen_margin_up', 'gen_margin_down', 'load_p', 'load_q', 'load_v', 'p_or', 'q_or', 'v_or', 'a_or', 'p_ex', 'q_ex', 'v_ex', 'a_ex', 'rho', 'line_status', 'timestep_overflow', 'topo_vect', 'time_before_cooldown_line', 'time_before_cooldown_sub', 'time_next_maintenance', 'duration_next_maintenance', 'target_dispatch', 'actual_dispatch', 'storage_charge', 'storage_power_target', 'storage_power', 'curtailment', 'curtailment_limit', 'curtailment_limit_effective', 'thermal_limit', 'is_alarm_illegal', 'time_since_last_alarm', 'last_alarm', 'attention_budget', 'was_alarm_used_after_game_over', 'max_step', 'active_alert', 'attack_under_alert', 'time_since_last_alert', 'alert_duration', 'total_number_of_alert', 'time_since_last_attack', 'was_alert_used_after_attack', 'theta_or', 'theta_ex', 'load_theta', 'gen_theta'), subtract=None, divide=None, functs=None)

This class allows to convert a grid2op observation space into a gym “Box” which is a regular Box in R^d.

It also allows to customize which part of the observation you want to use and offer capacity to center / reduce the data or to use more complex function from the observation.

Warning

Depending on the presence absence of gymnasium and gym packages this class might behave differently.

In grid2op we tried to maintain compatibility both with gymnasium (newest) and gym (legacy, no more maintained) RL packages. The behaviour is the following:

See Gymnasium vs Gym for more information

Note

A gymnasium Box is encoded as a numpy array.

Examples

If you simply want to use it you can do:

import grid2op
env_name = "l2rpn_case14_sandbox"  # or any other name
env = grid2op.make(env_name)

from grid2op.gym_compat import GymEnv, BoxGymObsSpace
gym_env = GymEnv(env)

gym_env.observation_space = BoxGymObsSpace(env.observation_space)

In this case it will extract all the features in all the observation (a detailed list is given in the documentation at Observation.

You can select the attribute you want to keep, for example:

gym_env.observation_space = BoxGymObsSpace(env.observation_space,
                                           attr_to_keep=['load_p', "gen_p", "rho])

You can also apply some basic transformation to the attribute of the observation before building the resulting gym observation (which in this case is a vector). This can be done with:

gym_env.observation_space = BoxGymObsSpace(env.observation_space,
                                           attr_to_keep=['load_p', "gen_p", "rho"],
                                           divide={"gen_p": env.gen_pmax},
                                           substract={"gen_p": 0.5 * env.gen_pmax})

In the above example, the resulting “gen_p” part of the vector will be given by the following formula: gym_obs = (grid2op_obs - substract) / divide.

Hint: you can use: divide being the standard deviation and subtract being the average of the attribute on a few episodes for example. This can be done with grid2op.utils.EpisodeStatistics for example.

Finally, you can also modify more the attribute of the observation and add it to your box. This can be done rather easily with the “functs” argument like:

gym_env.observation_space = BoxGymObsSpace(env.observation_space,
                                           attr_to_keep=["connectivity_matrix", "log_load"],
                                           functs={"connectivity_matrix":
                                                      (lambda grid2opobs: grid2opobs.connectivity_matrix().flatten(),
                                                       0., 1.0, None, None),
                                                   "log_load":
                                                    (lambda grid2opobs: np.log(grid2opobs.load_p),
                                                    None, 10., None, None)
                                                }
                                           )

In this case, “functs” should be a dictionary, the “keys” should be string (keys should also be present in the attr_to_keep list) and the values should count 5 elements (callable, low, high, shape, dtype) with:

  • callable a function taking as input a grid2op observation and returning a numpy array

  • low (optional) (put None if you don’t want to specify it, defaults to -np.inf) the lowest value your numpy array can take. It can be a single number or an array with the same shape as the return value of your function.

  • high (optional) (put None if you don’t want to specify it, defaults to np.inf) the highest value your numpy array can take. It can be a single number or an array with the same shape as the return value of your function.

  • shape (optional) (put None if you don’t want to specify it) the shape of the return value of your function. It should be a tuple (and not a single number). By default it is computed with by applying your function to an observation.

  • dtype (optional, put None if you don’t want to change it, defaults to np.float32) the type of the numpy array as output of your function.

Notes

The range of the values for “gen_p” / “prod_p” are not strictly env.gen_pmin and env.gen_pmax. This is due to the “approximation” when some redispatching is performed (the precision of the algorithm that computes the actual dispatch from the information it receives) and also because sometimes the losses of the grid are really different that the one anticipated in the “chronics” (yes env.gen_pmin and env.gen_pmax are not always ensured in grid2op)

grid2op.gym_compat.ContinuousToDiscreteConverter

alias of ContinuousToDiscreteConverterGymnasium

class grid2op.gym_compat.ContinuousToDiscreteConverterGymnasium(nb_bins, init_space=None)

TODO doc in progress

Some RL algorithms are particularly suited for dealing with discrete action space or observation space.

This “AttributeConverter” is responsible to convert continuous space to discrete space. The way it does it is by using bins. It uses np.linspace to compute the bins.

We recommend using an odd number of bins (eg 3, 7 or 9 for example).

Warning

Depending on the presence absence of gymnasium and gym packages this class might behave differently.

In grid2op we tried to maintain compatibility both with gymnasium (newest) and gym (legacy, no more maintained) RL packages. The behaviour is the following:

See Gymnasium vs Gym for more information

Examples

If nb_bins is 3 and the original input space is [-10, 10], then the split is the following:

  • 0 encodes all numbers in [-10, -3.33)

  • 1 encodes all numbers in [-3.33, 3.33)

  • 2 encode all numbers in [3.33, 10.]

And reciprocally, this action with :

  • 0 is understood as -5.0 (middle of the interval -10 / 0)

  • 1 is understood as 0.0 (middle of the interval represented by -10 / 10)

  • 2 is understood as 5.0 (middle of the interval represented by 0 / 10)

If nb_bins is 5 and the original input space is [-10, 10], then the split is the following:

  • 0 encodes all numbers in [-10, -6)

  • 1 encodes all numbers in [-6, -2)

  • 2 encode all numbers in [-2, 2)

  • 3 encode all numbers in [2, 6)

  • 4 encode all numbers in [6, 10]

And reciprocally, this action with :

  • 0 is understood as -6.6666…

  • 1 is understood as -3.333…

  • 2 is understood as 0.

  • 3 is understood as 3.333…

  • 4 is understood as 6.6666…

TODO add example of code on how to use this.

class grid2op.gym_compat.ContinuousToDiscreteConverterLegacyGym(nb_bins, init_space=None)

TODO doc in progress

Some RL algorithms are particularly suited for dealing with discrete action space or observation space.

This “AttributeConverter” is responsible to convert continuous space to discrete space. The way it does it is by using bins. It uses np.linspace to compute the bins.

We recommend using an odd number of bins (eg 3, 7 or 9 for example).

Warning

Depending on the presence absence of gymnasium and gym packages this class might behave differently.

In grid2op we tried to maintain compatibility both with gymnasium (newest) and gym (legacy, no more maintained) RL packages. The behaviour is the following:

See Gymnasium vs Gym for more information

Examples

If nb_bins is 3 and the original input space is [-10, 10], then the split is the following:

  • 0 encodes all numbers in [-10, -3.33)

  • 1 encodes all numbers in [-3.33, 3.33)

  • 2 encode all numbers in [3.33, 10.]

And reciprocally, this action with :

  • 0 is understood as -5.0 (middle of the interval -10 / 0)

  • 1 is understood as 0.0 (middle of the interval represented by -10 / 10)

  • 2 is understood as 5.0 (middle of the interval represented by 0 / 10)

If nb_bins is 5 and the original input space is [-10, 10], then the split is the following:

  • 0 encodes all numbers in [-10, -6)

  • 1 encodes all numbers in [-6, -2)

  • 2 encode all numbers in [-2, 2)

  • 3 encode all numbers in [2, 6)

  • 4 encode all numbers in [6, 10]

And reciprocally, this action with :

  • 0 is understood as -6.6666…

  • 1 is understood as -3.333…

  • 2 is understood as 0.

  • 3 is understood as 3.333…

  • 4 is understood as 6.6666…

TODO add example of code on how to use this.

grid2op.gym_compat.DiscreteActSpace

alias of DiscreteActSpaceGymnasium

class grid2op.gym_compat.DiscreteActSpaceGymnasium(grid2op_action_space: ActionSpace, attr_to_keep: Tuple[Literal['set_line_status'], Literal['set_line_status_simple'], Literal['change_line_status'], Literal['set_bus'], Literal['change_bus'], Literal['redispatch'], Literal['set_storage'], Literal['curtail'], Literal['curtail_mw']] | None = ('set_line_status', 'change_line_status', 'set_bus', 'change_bus', 'redispatch', 'set_storage', 'curtail'), nb_bins: Dict[Literal['redispatch', 'set_storage', 'curtail', 'curtail_mw'], int] | None = None, action_list=None)

TODO the documentation of this class is in progress.

This class allows to convert a grid2op action space into a gym “Discrete”. This means that the action are labeled, and instead of describing the action itself, you provide only its ID.

Let’s take an example of line disconnection. In the “standard” gym representation you need to:

import grid2op
import numpy as np
from grid2op.gym_compat import GymEnv

env_name = "l2rpn_case14_sandbox"  # or any other name
env = grid2op.make(env_name)
gym_env = GymEnv(env)

# now do an action
gym_act = {}
gym_act["set_bus"]  = np.zeros(env.n_line, dtype=np.int)
l_id = ... # the line you want to disconnect
gym_act["set_bus"][l_id] = -1
obs, reward, done, truncated, info = gym_env.step(gym_act)

This has the advantage to be as close as possible to raw grid2op. But the main drawback is that most of RL framework are not able to do this kind of modification easily. For discrete actions, what is often do is:

  1. enumerate all possible actions (say you have n different actions)

  2. assign a unique id to all actions (say from 0 to n-1)

  3. have a “policy” output a vector of size n with each component representing an action (eg vect[42] represents the score the policy assign to action 42)

Instead of having everyone doing the modifications “on its own” we developed the DiscreteActSpace that does exactly this, in a single line of code:

import grid2op
import numpy as np
from grid2op.gym_compat import GymEnv, DiscreteActSpace

env_name = "l2rpn_case14_sandbox"  # or any other name
env = grid2op.make(env_name)
gym_env = GymEnv(env)
gym_env.action_space = DiscreteActSpace(env.action_space,
                                        attr_to_keep=["set_bus",
                                                      "set_line_status",
                                                      # or anything else
                                                     ]
                                        )

# do action with ID 42
gym_act = 42
obs, reward, done, truncated, info = gym_env.step(gym_act)
# to know what the action did, you can
# print(gym_env.action_space.from_gym(gym_act))

It is related to the MultiDiscreteActSpace but compared to this other representation, it does not allow to do “multiple actions”. Typically, if you use the snippets below:

import grid2op
env_name = "l2rpn_case14_sandbox"  # or any other name
env = grid2op.make(env_name)

from grid2op.gym_compat import GymEnv, MultiDiscreteActSpace, DiscreteActSpace
gym_env1 = GymEnv(env)
gym_env2 = GymEnv(env)

gym_env1.action_space = MultiDiscreteActSpace(env.action_space,
                                              attr_to_keep=['redispatch', "curtail", "one_sub_set"])
gym_env2.action_space = DiscreteActSpace(env.action_space,
                                         attr_to_keep=['redispatch', "curtail", "set_bus"])

Then at each step, gym_env1 will allow to perform a redispatching action (on any number of generators), a curtailment action (on any number of generators) __**AND**__ changing the topology at one substation. But at each steps, the agent should predicts lots of “number”.

On the other hand, at each step, the agent for gym_env2 will have to predict a single integer (which is usually the case in most RL environment) but it this action will affect redispatching on a single generator, perform curtailment on a single generator __**OR**__ changing the topology at one substation. But at each steps, the agent should predicts only one “number”.

The action set is then largely constrained compared to the MultiDiscreteActSpace

Note

This class is really closely related to the grid2op.Converter.IdToAct. It basically “maps” this “IdToAct” into a type of gym space, which, in this case, will be a Discrete one.

Note

By default, the “do nothing” action is encoded by the integer ‘0’.

Warning

Depending on the presence absence of gymnasium and gym packages this class might behave differently.

In grid2op we tried to maintain compatibility both with gymnasium (newest) and gym (legacy, no more maintained) RL packages. The behaviour is the following:

See Gymnasium vs Gym for more information

Examples

We recommend to use it like:

import grid2op
env_name = "l2rpn_case14_sandbox"  # or any other name
env = grid2op.make(env_name)

from grid2op.gym_compat import GymEnv, DiscreteActSpace
gym_env = GymEnv(env)

gym_env.observation_space = DiscreteActSpace(env.observation_space,
                                             attr_to_keep=['redispatch', "curtail", "set_bus"])

The possible attribute you can provide in the “attr_to_keep” are:

  • “set_line_status”

  • “set_line_status_simple” (grid2op >= 1.6.6) : set line status adds 5 actions per powerlines:

    1. disconnect it

    2. connect origin side to busbar 1 and extermity side to busbar 1

    3. connect origin side to busbar 1 and extermity side to busbar 2

    4. connect origin side to busbar 2 and extermity side to busbar 1

    5. connect origin side to busbar 2 and extermity side to busbar 2

    This is “over complex” for most use case where you just want to “connect it” or “disconnect it”. If you want the simplest version, just use “set_line_status_simple”.

  • “change_line_status”

  • “set_bus”: corresponds to changing the topology using the “set_bus” (equivalent to the “one_sub_set” keyword in the “attr_to_keep” of the MultiDiscreteActSpace)

  • “change_bus”: corresponds to changing the topology using the “change_bus” (equivalent to the “one_sub_change” keyword in the “attr_to_keep” of the MultiDiscreteActSpace)

  • “redispatch”

  • “set_storage”

  • “curtail”

  • “curtail_mw” (same effect as “curtail”)

If you do not want (each time) to build all the actions from the action space, but would rather save the actions you find the most interesting and then reload them, you can, for example:

import grid2op
from grid2op.gym_compat import GymEnv, DiscreteActSpace
env_name = "l2rpn_case14_sandbox"  # or any other name
env = grid2op.make(env_name)

gym_env = GymEnv(env)
action_list = ... # a list of action, that can be processed by
# IdToAct.init_converter (all_actions): see
# https://grid2op.readthedocs.io/en/latest/converter.html#grid2op.Converter.IdToAct.init_converter
gym_env.observation_space = DiscreteActSpace(env.observation_space,
                                             action_list=action_list)

Note

This last version (providing explicitly the actions you want to keep and their ID) is much (much) safer and reproducible. Indeed, the actions usable by your agent will be the same (and in the same order) regardless of the grid2op version, of the person using it, of pretty much everything.

It might not be consistent (between different grid2op versions) if the actions are built from scratch (for example, depending on the grid2op version other types of actions can be made, such as curtailment, or actions on storage units) like it’s the case with the key-words (eg “set_bus”) you pass as argument in the attr_to_keep

class grid2op.gym_compat.DiscreteActSpaceLegacyGym(grid2op_action_space: ActionSpace, attr_to_keep: Tuple[Literal['set_line_status'], Literal['set_line_status_simple'], Literal['change_line_status'], Literal['set_bus'], Literal['change_bus'], Literal['redispatch'], Literal['set_storage'], Literal['curtail'], Literal['curtail_mw']] | None = ('set_line_status', 'change_line_status', 'set_bus', 'change_bus', 'redispatch', 'set_storage', 'curtail'), nb_bins: Dict[Literal['redispatch', 'set_storage', 'curtail', 'curtail_mw'], int] | None = None, action_list=None)

TODO the documentation of this class is in progress.

This class allows to convert a grid2op action space into a gym “Discrete”. This means that the action are labeled, and instead of describing the action itself, you provide only its ID.

Let’s take an example of line disconnection. In the “standard” gym representation you need to:

import grid2op
import numpy as np
from grid2op.gym_compat import GymEnv

env_name = "l2rpn_case14_sandbox"  # or any other name
env = grid2op.make(env_name)
gym_env = GymEnv(env)

# now do an action
gym_act = {}
gym_act["set_bus"]  = np.zeros(env.n_line, dtype=np.int)
l_id = ... # the line you want to disconnect
gym_act["set_bus"][l_id] = -1
obs, reward, done, truncated, info = gym_env.step(gym_act)

This has the advantage to be as close as possible to raw grid2op. But the main drawback is that most of RL framework are not able to do this kind of modification easily. For discrete actions, what is often do is:

  1. enumerate all possible actions (say you have n different actions)

  2. assign a unique id to all actions (say from 0 to n-1)

  3. have a “policy” output a vector of size n with each component representing an action (eg vect[42] represents the score the policy assign to action 42)

Instead of having everyone doing the modifications “on its own” we developed the DiscreteActSpace that does exactly this, in a single line of code:

import grid2op
import numpy as np
from grid2op.gym_compat import GymEnv, DiscreteActSpace

env_name = "l2rpn_case14_sandbox"  # or any other name
env = grid2op.make(env_name)
gym_env = GymEnv(env)
gym_env.action_space = DiscreteActSpace(env.action_space,
                                        attr_to_keep=["set_bus",
                                                      "set_line_status",
                                                      # or anything else
                                                     ]
                                        )

# do action with ID 42
gym_act = 42
obs, reward, done, truncated, info = gym_env.step(gym_act)
# to know what the action did, you can
# print(gym_env.action_space.from_gym(gym_act))

It is related to the MultiDiscreteActSpace but compared to this other representation, it does not allow to do “multiple actions”. Typically, if you use the snippets below:

import grid2op
env_name = "l2rpn_case14_sandbox"  # or any other name
env = grid2op.make(env_name)

from grid2op.gym_compat import GymEnv, MultiDiscreteActSpace, DiscreteActSpace
gym_env1 = GymEnv(env)
gym_env2 = GymEnv(env)

gym_env1.action_space = MultiDiscreteActSpace(env.action_space,
                                              attr_to_keep=['redispatch', "curtail", "one_sub_set"])
gym_env2.action_space = DiscreteActSpace(env.action_space,
                                         attr_to_keep=['redispatch', "curtail", "set_bus"])

Then at each step, gym_env1 will allow to perform a redispatching action (on any number of generators), a curtailment action (on any number of generators) __**AND**__ changing the topology at one substation. But at each steps, the agent should predicts lots of “number”.

On the other hand, at each step, the agent for gym_env2 will have to predict a single integer (which is usually the case in most RL environment) but it this action will affect redispatching on a single generator, perform curtailment on a single generator __**OR**__ changing the topology at one substation. But at each steps, the agent should predicts only one “number”.

The action set is then largely constrained compared to the MultiDiscreteActSpace

Note

This class is really closely related to the grid2op.Converter.IdToAct. It basically “maps” this “IdToAct” into a type of gym space, which, in this case, will be a Discrete one.

Note

By default, the “do nothing” action is encoded by the integer ‘0’.

Warning

Depending on the presence absence of gymnasium and gym packages this class might behave differently.

In grid2op we tried to maintain compatibility both with gymnasium (newest) and gym (legacy, no more maintained) RL packages. The behaviour is the following:

See Gymnasium vs Gym for more information

Examples

We recommend to use it like:

import grid2op
env_name = "l2rpn_case14_sandbox"  # or any other name
env = grid2op.make(env_name)

from grid2op.gym_compat import GymEnv, DiscreteActSpace
gym_env = GymEnv(env)

gym_env.observation_space = DiscreteActSpace(env.observation_space,
                                             attr_to_keep=['redispatch', "curtail", "set_bus"])

The possible attribute you can provide in the “attr_to_keep” are:

  • “set_line_status”

  • “set_line_status_simple” (grid2op >= 1.6.6) : set line status adds 5 actions per powerlines:

    1. disconnect it

    2. connect origin side to busbar 1 and extermity side to busbar 1

    3. connect origin side to busbar 1 and extermity side to busbar 2

    4. connect origin side to busbar 2 and extermity side to busbar 1

    5. connect origin side to busbar 2 and extermity side to busbar 2

    This is “over complex” for most use case where you just want to “connect it” or “disconnect it”. If you want the simplest version, just use “set_line_status_simple”.

  • “change_line_status”

  • “set_bus”: corresponds to changing the topology using the “set_bus” (equivalent to the “one_sub_set” keyword in the “attr_to_keep” of the MultiDiscreteActSpace)

  • “change_bus”: corresponds to changing the topology using the “change_bus” (equivalent to the “one_sub_change” keyword in the “attr_to_keep” of the MultiDiscreteActSpace)

  • “redispatch”

  • “set_storage”

  • “curtail”

  • “curtail_mw” (same effect as “curtail”)

If you do not want (each time) to build all the actions from the action space, but would rather save the actions you find the most interesting and then reload them, you can, for example:

import grid2op
from grid2op.gym_compat import GymEnv, DiscreteActSpace
env_name = "l2rpn_case14_sandbox"  # or any other name
env = grid2op.make(env_name)

gym_env = GymEnv(env)
action_list = ... # a list of action, that can be processed by
# IdToAct.init_converter (all_actions): see
# https://grid2op.readthedocs.io/en/latest/converter.html#grid2op.Converter.IdToAct.init_converter
gym_env.observation_space = DiscreteActSpace(env.observation_space,
                                             action_list=action_list)

Note

This last version (providing explicitly the actions you want to keep and their ID) is much (much) safer and reproducible. Indeed, the actions usable by your agent will be the same (and in the same order) regardless of the grid2op version, of the person using it, of pretty much everything.

It might not be consistent (between different grid2op versions) if the actions are built from scratch (for example, depending on the grid2op version other types of actions can be made, such as curtailment, or actions on storage units) like it’s the case with the key-words (eg “set_bus”) you pass as argument in the attr_to_keep

grid2op.gym_compat.GymActionSpace

alias of GymnasiumActionSpace

grid2op.gym_compat.GymEnv

alias of GymnasiumEnv Methods:

reset(*[, seed, options])

Resets the environment to an initial internal state, returning an initial observation and info.

step(action)

Run one timestep of the environment's dynamics using the agent actions.

class grid2op.gym_compat.GymEnv_Legacy(env_init: Environment, shuffle_chronics: bool | None = True, render_mode: Literal['rgb_array'] = 'rgb_array')[source]

fully implements the openAI gym API by using the GymActionSpace and GymObservationSpace for compliance with openAI gym.

They can handle action_space_converter or observation_space converter to change the representation of data that will be fed to the agent. #TODO

Warning

The gym package has some breaking API change since its version 0.26. Depending on the version installed, we attempted, in grid2op, to maintain compatibility both with former version and later one. This makes this class behave differently depending on the version of gym you have installed !

The main changes involve the functions env.step and env.reset

If you want to use the same version of the GymEnv regardless of the gym / gymnasium version installed you can use:

Warning

Depending on the presence absence of gymnasium and gym packages this class might behave differently.

In grid2op we tried to maintain compatibility both with gymnasium (newest) and gym (legacy, no more maintained) RL packages. The behaviour is the following:

See Gymnasium vs Gym for more information

Notes

The environment passed as input is copied. It is not modified by this “gym environment”

Examples

This can be used like:

import grid2op
from grid2op.gym_compat import GymEnv

env_name = "l2rpn_case14_sandbox"  # or any other name
env = grid2op.make(env_name)
gym_env = GymEnv(env)  # is a gym environment properly inheriting from gym.Env !

There are a few difference between “raw” grid2op environment and gymnasium environments.

One of the major difference is that, to our knowledge, gymnasium does not support the simulate feature (which allows an agent to test the impact of a given action on the grid without having to perform a step see Model Based / Planning methods for more information) [NB if you know or better are developping some “model based RL library” let us know !]

Another difference is in the way to do some actions. In grid2op, actions are a dedicated class and can be made with an action_space and a dictionary, or using the properties of the action class.

In gym, there are no specific representations of the action class. More precisely, for each action type (MultiDiscreteActSpace, DiscreteActSpace, BoxGymActSpace or GymActionSpace) there is a way to encode it. For example, by default (GymActionSpace) an action is represented through an OrderedDict (from collection import OrderedDict)

Methods:

reset(*args, **kwargs)

Resets the environment to an initial state and returns the initial observation.

step(action)

Run one timestep of the environment's dynamics.

reset(*args, **kwargs) ObsType[source]

Resets the environment to an initial state and returns the initial observation.

This method can reset the environment’s random number generator(s) if seed is an integer or if the environment has not yet initialized a random number generator. If the environment already has a random number generator and reset() is called with seed=None, the RNG should not be reset. Moreover, reset() should (in the typical use case) be called with an integer seed right after initialization and then never again.

Parameters:
  • seed (optional int) – The seed that is used to initialize the environment’s PRNG. If the environment does not already have a PRNG and seed=None (the default option) is passed, a seed will be chosen from some source of entropy (e.g. timestamp or /dev/urandom). However, if the environment already has a PRNG and seed=None is passed, the PRNG will not be reset. If you pass an integer, the PRNG will be reset even if it already exists. Usually, you want to pass an integer right after the environment has been initialized and then never again. Please refer to the minimal example above to see this paradigm in action.

  • options (optional dict) – Additional information to specify how the environment is reset (optional, depending on the specific environment)

Returns:

Observation of the initial state. This will be an element of observation_space

(typically a numpy array) and is analogous to the observation returned by step().

info (dictionary): This dictionary contains auxiliary information complementing observation. It should be analogous to

the info returned by step().

Return type:

observation (object)

step(action: ActType) Tuple[ObsType, float, bool, Dict[Literal['disc_lines', 'is_illegal', 'is_ambiguous', 'is_dispatching_illegal', 'is_illegal_reco', 'reason_alarm_illegal', 'reason_alert_illegal', 'opponent_attack_line', 'opponent_attack_sub', 'exception', 'detailed_infos_for_cascading_failures', 'rewards', 'time_series_id'], Any]][source]

Run one timestep of the environment’s dynamics.

When end of episode is reached, you are responsible for calling reset() to reset this environment’s state. Accepts an action and returns either a tuple (observation, reward, terminated, truncated, info).

Parameters:

action (ActType) – an action provided by the agent

Returns:

this will be an element of the environment’s observation_space.

This may, for instance, be a numpy array containing the positions and velocities of certain objects.

reward (float): The amount of reward returned as a result of taking the action. terminated (bool): whether a terminal state (as defined under the MDP of the task) is reached.

In this case further step() calls could return undefined results.

truncated (bool): whether a truncation condition outside the scope of the MDP is satisfied.

Typically a timelimit, but could also be used to indicate agent physically going out of bounds. Can be used to end the episode prematurely before a terminal state is reached.

info (dictionary): info contains auxiliary diagnostic information (helpful for debugging, learning, and logging).

This might, for instance, contain: metrics that describe the agent’s performance state, variables that are hidden from observations, or individual reward terms that are combined to produce the total reward. It also can contain information that distinguishes truncation and termination, however this is deprecated in favour of returning two booleans, and will be removed in a future version.

(deprecated) done (bool): A boolean value for if the episode has ended, in which case further step() calls will return undefined results.

A done signal may be emitted for different reasons: Maybe the task underlying the environment was solved successfully, a certain timelimit was exceeded, or the physics simulation has entered an invalid state.

Return type:

observation (object)

class grid2op.gym_compat.GymEnv_Modern(env_init: Environment, shuffle_chronics: bool | None = True, render_mode: Literal['rgb_array'] = 'rgb_array')[source]

fully implements the openAI gym API by using the GymActionSpace and GymObservationSpace for compliance with openAI gym.

They can handle action_space_converter or observation_space converter to change the representation of data that will be fed to the agent. #TODO

Warning

The gym package has some breaking API change since its version 0.26. Depending on the version installed, we attempted, in grid2op, to maintain compatibility both with former version and later one. This makes this class behave differently depending on the version of gym you have installed !

The main changes involve the functions env.step and env.reset

If you want to use the same version of the GymEnv regardless of the gym / gymnasium version installed you can use:

Warning

Depending on the presence absence of gymnasium and gym packages this class might behave differently.

In grid2op we tried to maintain compatibility both with gymnasium (newest) and gym (legacy, no more maintained) RL packages. The behaviour is the following:

See Gymnasium vs Gym for more information

Notes

The environment passed as input is copied. It is not modified by this “gym environment”

Examples

This can be used like:

import grid2op
from grid2op.gym_compat import GymEnv

env_name = "l2rpn_case14_sandbox"  # or any other name
env = grid2op.make(env_name)
gym_env = GymEnv(env)  # is a gym environment properly inheriting from gym.Env !

There are a few difference between “raw” grid2op environment and gymnasium environments.

One of the major difference is that, to our knowledge, gymnasium does not support the simulate feature (which allows an agent to test the impact of a given action on the grid without having to perform a step see Model Based / Planning methods for more information) [NB if you know or better are developping some “model based RL library” let us know !]

Another difference is in the way to do some actions. In grid2op, actions are a dedicated class and can be made with an action_space and a dictionary, or using the properties of the action class.

In gym, there are no specific representations of the action class. More precisely, for each action type (MultiDiscreteActSpace, DiscreteActSpace, BoxGymActSpace or GymActionSpace) there is a way to encode it. For example, by default (GymActionSpace) an action is represented through an OrderedDict (from collection import OrderedDict)

Methods:

reset(*[, seed, options])

Resets the environment to an initial state and returns the initial observation.

step(action)

Run one timestep of the environment's dynamics.

reset(*, seed: int | None = None, options: Dict[str | Literal['time serie id'], int | str] | None = None) Tuple[ObsType, Dict[Literal['time serie id', 'seed', 'grid2op_env_seed', 'underlying_env_seeds'], Any]][source]

Resets the environment to an initial state and returns the initial observation.

This method can reset the environment’s random number generator(s) if seed is an integer or if the environment has not yet initialized a random number generator. If the environment already has a random number generator and reset() is called with seed=None, the RNG should not be reset. Moreover, reset() should (in the typical use case) be called with an integer seed right after initialization and then never again.

Parameters:
  • seed (optional int) – The seed that is used to initialize the environment’s PRNG. If the environment does not already have a PRNG and seed=None (the default option) is passed, a seed will be chosen from some source of entropy (e.g. timestamp or /dev/urandom). However, if the environment already has a PRNG and seed=None is passed, the PRNG will not be reset. If you pass an integer, the PRNG will be reset even if it already exists. Usually, you want to pass an integer right after the environment has been initialized and then never again. Please refer to the minimal example above to see this paradigm in action.

  • options (optional dict) – Additional information to specify how the environment is reset (optional, depending on the specific environment)

Returns:

Observation of the initial state. This will be an element of observation_space

(typically a numpy array) and is analogous to the observation returned by step().

info (dictionary): This dictionary contains auxiliary information complementing observation. It should be analogous to

the info returned by step().

Return type:

observation (object)

step(action: ActType) Tuple[ObsType, float, bool, bool, Dict[Literal['disc_lines', 'is_illegal', 'is_ambiguous', 'is_dispatching_illegal', 'is_illegal_reco', 'reason_alarm_illegal', 'reason_alert_illegal', 'opponent_attack_line', 'opponent_attack_sub', 'exception', 'detailed_infos_for_cascading_failures', 'rewards', 'time_series_id'], Any]][source]

Run one timestep of the environment’s dynamics.

When end of episode is reached, you are responsible for calling reset() to reset this environment’s state. Accepts an action and returns either a tuple (observation, reward, terminated, truncated, info).

Parameters:

action (ActType) – an action provided by the agent

Returns:

this will be an element of the environment’s observation_space.

This may, for instance, be a numpy array containing the positions and velocities of certain objects.

reward (float): The amount of reward returned as a result of taking the action. terminated (bool): whether a terminal state (as defined under the MDP of the task) is reached.

In this case further step() calls could return undefined results.

truncated (bool): whether a truncation condition outside the scope of the MDP is satisfied.

Typically a timelimit, but could also be used to indicate agent physically going out of bounds. Can be used to end the episode prematurely before a terminal state is reached.

info (dictionary): info contains auxiliary diagnostic information (helpful for debugging, learning, and logging).

This might, for instance, contain: metrics that describe the agent’s performance state, variables that are hidden from observations, or individual reward terms that are combined to produce the total reward. It also can contain information that distinguishes truncation and termination, however this is deprecated in favour of returning two booleans, and will be removed in a future version.

(deprecated) done (bool): A boolean value for if the episode has ended, in which case further step() calls will return undefined results.

A done signal may be emitted for different reasons: Maybe the task underlying the environment was solved successfully, a certain timelimit was exceeded, or the physics simulation has entered an invalid state.

Return type:

observation (object)

grid2op.gym_compat.GymObservationSpace

alias of GymnasiumObservationSpace

class grid2op.gym_compat.GymnasiumActionSpace(env, converter=None, dict_variables=None)

This class enables the conversion of the action space into a gym “space”.

Resulting action space will be a gym.spaces.Dict.

NB it is NOT recommended to use the sample of the gym action space. Please use the sampling ( if availabe) of the original action space instead [if not available this means there is no implemented way to generate reliable random action]

Note that gym space converted with this class should be seeded independently. It is NOT seeded when calling grid2op.Environment.Environment.seed().

Warning

Depending on the presence absence of gymnasium and gym packages this class might behave differently.

In grid2op we tried to maintain compatibility both with gymnasium (newest) and gym (legacy, no more maintained) RL packages. The behaviour is the following:

See Gymnasium vs Gym for more information

Note

A gymnasium Dict is encoded as a OrderedDict (from collection import OrderedDict) see the example section for more information.

Examples

For the “l2rpn_case14_sandbox” environment, a code using BoxGymActSpace can look something like (if you want to build action “by hands”):

import grid2op
from grid2op.gym_compat import GymEnv
import numpy as np
env_name = "l2rpn_case14_sandbox"

env = grid2op.make(env_name)
gym_env =  GymEnv(env)

obs = gym_env.reset()  # obs will be an OrderedDict (default, but you can customize it)

# is equivalent to "do nothing"
act = {}
obs, reward, done, truncated, info = gym_env.step(act)

# you can also do a random action:
act = gym_env.action_space.sample()
print(gym_env.action_space.from_gym(act))
obs, reward, done, truncated, info = gym_env.step(act)

# you can chose the action you want to do (say "redispatch" for example)
# here a random redispatch action
act = {}
attr_nm = "redispatch"
act[attr_nm] = np.random.uniform(high=gym_env.action_space.spaces[attr_nm].low,
                                low=gym_env.action_space.spaces[attr_nm].high,
                                size=env.n_gen)
print(gym_env.action_space.from_gym(act))
obs, reward, done, truncated, info = gym_env.step(act)
class grid2op.gym_compat.GymnasiumEnv(env_init: Environment, shuffle_chronics: bool | None = True, render_mode: Literal['rgb_array'] = 'rgb_array')[source]

fully implements the openAI gym API by using the GymActionSpace and GymObservationSpace for compliance with openAI gym.

They can handle action_space_converter or observation_space converter to change the representation of data that will be fed to the agent. #TODO

Warning

The gym package has some breaking API change since its version 0.26. Depending on the version installed, we attempted, in grid2op, to maintain compatibility both with former version and later one. This makes this class behave differently depending on the version of gym you have installed !

The main changes involve the functions env.step and env.reset

If you want to use the same version of the GymEnv regardless of the gym / gymnasium version installed you can use:

Warning

Depending on the presence absence of gymnasium and gym packages this class might behave differently.

In grid2op we tried to maintain compatibility both with gymnasium (newest) and gym (legacy, no more maintained) RL packages. The behaviour is the following:

See Gymnasium vs Gym for more information

Notes

The environment passed as input is copied. It is not modified by this “gym environment”

Examples

This can be used like:

import grid2op
from grid2op.gym_compat import GymEnv

env_name = "l2rpn_case14_sandbox"  # or any other name
env = grid2op.make(env_name)
gym_env = GymEnv(env)  # is a gym environment properly inheriting from gym.Env !

There are a few difference between “raw” grid2op environment and gymnasium environments.

One of the major difference is that, to our knowledge, gymnasium does not support the simulate feature (which allows an agent to test the impact of a given action on the grid without having to perform a step see Model Based / Planning methods for more information) [NB if you know or better are developping some “model based RL library” let us know !]

Another difference is in the way to do some actions. In grid2op, actions are a dedicated class and can be made with an action_space and a dictionary, or using the properties of the action class.

In gym, there are no specific representations of the action class. More precisely, for each action type (MultiDiscreteActSpace, DiscreteActSpace, BoxGymActSpace or GymActionSpace) there is a way to encode it. For example, by default (GymActionSpace) an action is represented through an OrderedDict (from collection import OrderedDict)

Methods:

reset(*[, seed, options])

Resets the environment to an initial internal state, returning an initial observation and info.

step(action)

Run one timestep of the environment's dynamics using the agent actions.

reset(*, seed: int | None = None, options: Dict[str | Literal['time serie id'], int | str] | None = None) Tuple[ObsType, Dict[Literal['time serie id', 'seed', 'grid2op_env_seed', 'underlying_env_seeds'], Any]][source]

Resets the environment to an initial internal state, returning an initial observation and info.

This method generates a new starting state often with some randomness to ensure that the agent explores the state space and learns a generalised policy about the environment. This randomness can be controlled with the seed parameter otherwise if the environment already has a random number generator and reset() is called with seed=None, the RNG is not reset.

Therefore, reset() should (in the typical use case) be called with a seed right after initialization and then never again.

For Custom environments, the first line of reset() should be super().reset(seed=seed) which implements the seeding correctly.

Changed in version v0.25: The return_info parameter was removed and now info is expected to be returned.

Parameters:
  • seed (optional int) – The seed that is used to initialize the environment’s PRNG (np_random). If the environment does not already have a PRNG and seed=None (the default option) is passed, a seed will be chosen from some source of entropy (e.g. timestamp or /dev/urandom). However, if the environment already has a PRNG and seed=None is passed, the PRNG will not be reset. If you pass an integer, the PRNG will be reset even if it already exists. Usually, you want to pass an integer right after the environment has been initialized and then never again. Please refer to the minimal example above to see this paradigm in action.

  • options (optional dict) – Additional information to specify how the environment is reset (optional, depending on the specific environment)

Returns:

Observation of the initial state. This will be an element of observation_space

(typically a numpy array) and is analogous to the observation returned by step().

info (dictionary): This dictionary contains auxiliary information complementing observation. It should be analogous to

the info returned by step().

Return type:

observation (ObsType)

step(action: ActType) Tuple[ObsType, float, bool, bool, Dict[Literal['disc_lines', 'is_illegal', 'is_ambiguous', 'is_dispatching_illegal', 'is_illegal_reco', 'reason_alarm_illegal', 'reason_alert_illegal', 'opponent_attack_line', 'opponent_attack_sub', 'exception', 'detailed_infos_for_cascading_failures', 'rewards', 'time_series_id'], Any]][source]

Run one timestep of the environment’s dynamics using the agent actions.

When the end of an episode is reached (terminated or truncated), it is necessary to call reset() to reset this environment’s state for the next episode.

Changed in version 0.26: The Step API was changed removing done in favor of terminated and truncated to make it clearer to users when the environment had terminated or truncated which is critical for reinforcement learning bootstrapping algorithms.

Parameters:

action (ActType) – an action provided by the agent to update the environment state.

Returns:

An element of the environment’s observation_space as the next observation due to the agent actions.

An example is a numpy array containing the positions and velocities of the pole in CartPole.

reward (SupportsFloat): The reward as a result of taking the action. terminated (bool): Whether the agent reaches the terminal state (as defined under the MDP of the task)

which can be positive or negative. An example is reaching the goal state or moving into the lava from the Sutton and Barton, Gridworld. If true, the user needs to call reset().

truncated (bool): Whether the truncation condition outside the scope of the MDP is satisfied.

Typically, this is a timelimit, but could also be used to indicate an agent physically going out of bounds. Can be used to end the episode prematurely before a terminal state is reached. If true, the user needs to call reset().

info (dict): Contains auxiliary diagnostic information (helpful for debugging, learning, and logging).

This might, for instance, contain: metrics that describe the agent’s performance state, variables that are hidden from observations, or individual reward terms that are combined to produce the total reward. In OpenAI Gym <v26, it contains “TimeLimit.truncated” to distinguish truncation and termination, however this is deprecated in favour of returning terminated and truncated variables.

done (bool): (Deprecated) A boolean value for if the episode has ended, in which case further step() calls will

return undefined results. This was removed in OpenAI Gym v26 in favor of terminated and truncated attributes. A done signal may be emitted for different reasons: Maybe the task underlying the environment was solved successfully, a certain timelimit was exceeded, or the physics simulation has entered an invalid state.

Return type:

observation (ObsType)

class grid2op.gym_compat.GymnasiumObservationSpace(env, dict_variables=None)

TODO explain gym / gymnasium

This class allows to transform the observation space into a gym space.

Gym space will be a gym.spaces.Dict with the keys being the different attributes of the grid2op observation. All attributes are used.

Note that gym space converted with this class should be seeded independently. It is NOT seeded when calling grid2op.Environment.Environment.seed().

Warning

Depending on the presence absence of gymnasium and gym packages this class might behave differently.

In grid2op we tried to maintain compatibility both with gymnasium (newest) and gym (legacy, no more maintained) RL packages. The behaviour is the following:

See Gymnasium vs Gym for more information

Examples

Converting an observation space is fairly straightforward:

import grid2op
from grid2op.Converter import GymObservationSpace
env = grid2op.make("l2rpn_case14_sandbox")

gym_observation_space = GymObservationSpace(env.observation_space)
# and now gym_observation_space is a `gym.spaces.Dict` representing the observation space

# you can "convert" the grid2op observation to / from this space with:

grid2op_obs = env.reset()
same_gym_obs = gym_observation_space.to_gym(grid2op_obs)

# the conversion from gym_obs to grid2op obs is feasible, but i don't imagine
# a situation where it is useful. And especially, you will not be able to
# use "obs.simulate" for the observation converted back from this gym action.

Notes

The range of the values for “gen_p” / “prod_p” are not strictly env.gen_pmin and env.gen_pmax. This is due to the “approximation” when some redispatching is performed (the precision of the algorithm that computes the actual dispatch from the information it receives) and also because sometimes the losses of the grid are really different that the one anticipated in the “chronics” (yes env.gen_pmin and env.gen_pmax are not always ensured in grid2op)

class grid2op.gym_compat.LegacyGymActionSpace(env, converter=None, dict_variables=None)

This class enables the conversion of the action space into a gym “space”.

Resulting action space will be a gym.spaces.Dict.

NB it is NOT recommended to use the sample of the gym action space. Please use the sampling ( if availabe) of the original action space instead [if not available this means there is no implemented way to generate reliable random action]

Note that gym space converted with this class should be seeded independently. It is NOT seeded when calling grid2op.Environment.Environment.seed().

Warning

Depending on the presence absence of gymnasium and gym packages this class might behave differently.

In grid2op we tried to maintain compatibility both with gymnasium (newest) and gym (legacy, no more maintained) RL packages. The behaviour is the following:

See Gymnasium vs Gym for more information

Note

A gymnasium Dict is encoded as a OrderedDict (from collection import OrderedDict) see the example section for more information.

Examples

For the “l2rpn_case14_sandbox” environment, a code using BoxGymActSpace can look something like (if you want to build action “by hands”):

import grid2op
from grid2op.gym_compat import GymEnv
import numpy as np
env_name = "l2rpn_case14_sandbox"

env = grid2op.make(env_name)
gym_env =  GymEnv(env)

obs = gym_env.reset()  # obs will be an OrderedDict (default, but you can customize it)

# is equivalent to "do nothing"
act = {}
obs, reward, done, truncated, info = gym_env.step(act)

# you can also do a random action:
act = gym_env.action_space.sample()
print(gym_env.action_space.from_gym(act))
obs, reward, done, truncated, info = gym_env.step(act)

# you can chose the action you want to do (say "redispatch" for example)
# here a random redispatch action
act = {}
attr_nm = "redispatch"
act[attr_nm] = np.random.uniform(high=gym_env.action_space.spaces[attr_nm].low,
                                low=gym_env.action_space.spaces[attr_nm].high,
                                size=env.n_gen)
print(gym_env.action_space.from_gym(act))
obs, reward, done, truncated, info = gym_env.step(act)
class grid2op.gym_compat.LegacyGymObservationSpace(env, dict_variables=None)

TODO explain gym / gymnasium

This class allows to transform the observation space into a gym space.

Gym space will be a gym.spaces.Dict with the keys being the different attributes of the grid2op observation. All attributes are used.

Note that gym space converted with this class should be seeded independently. It is NOT seeded when calling grid2op.Environment.Environment.seed().

Warning

Depending on the presence absence of gymnasium and gym packages this class might behave differently.

In grid2op we tried to maintain compatibility both with gymnasium (newest) and gym (legacy, no more maintained) RL packages. The behaviour is the following:

See Gymnasium vs Gym for more information

Examples

Converting an observation space is fairly straightforward:

import grid2op
from grid2op.Converter import GymObservationSpace
env = grid2op.make("l2rpn_case14_sandbox")

gym_observation_space = GymObservationSpace(env.observation_space)
# and now gym_observation_space is a `gym.spaces.Dict` representing the observation space

# you can "convert" the grid2op observation to / from this space with:

grid2op_obs = env.reset()
same_gym_obs = gym_observation_space.to_gym(grid2op_obs)

# the conversion from gym_obs to grid2op obs is feasible, but i don't imagine
# a situation where it is useful. And especially, you will not be able to
# use "obs.simulate" for the observation converted back from this gym action.

Notes

The range of the values for “gen_p” / “prod_p” are not strictly env.gen_pmin and env.gen_pmax. This is due to the “approximation” when some redispatching is performed (the precision of the algorithm that computes the actual dispatch from the information it receives) and also because sometimes the losses of the grid are really different that the one anticipated in the “chronics” (yes env.gen_pmin and env.gen_pmax are not always ensured in grid2op)

grid2op.gym_compat.MultiDiscreteActSpace

alias of MultiDiscreteActSpaceGymnasium

class grid2op.gym_compat.MultiDiscreteActSpaceGymnasium(grid2op_action_space: ActionSpace, attr_to_keep: Tuple[Literal['set_line_status'], Literal['change_line_status'], Literal['set_bus'], Literal['sub_set_bus'], Literal['one_sub_set'], Literal['change_bus'], Literal['sub_change_bus'], Literal['one_sub_change'], Literal['redispatch'], Literal['set_storage'], Literal['curtail'], Literal['curtail_mw'], Literal['one_line_set'], Literal['one_line_change']] | None = ('set_line_status', 'change_line_status', 'set_bus', 'change_bus', 'redispatch', 'set_storage', 'curtail', 'raise_alarm', 'raise_alert'), nb_bins: Dict[Literal['redispatch', 'set_storage', 'curtail', 'curtail_mw'], int] | None = None)

This class allows to convert a grid2op action space into a gym “MultiDiscrete”. This means that the action are labeled, and instead of describing the action itself, you provide only its ID.

Note

This action space is particularly suited for represented discrete actions.

It is possible to represent continuous actions with it. In that case, the continuous actions are “binarized” thanks to the ContinuousToDiscreteConverter. Feel free to consult its documentation for more information.

In this case it will extract all the features in all the action with:

  • “set_line_status”: n_line dimensions, each containing 3 choices “DISCONNECT”, “DONT AFFECT”, “FORCE CONNECTION” and affecting the powerline status (connected / disconnected)

  • “change_line_status”: n_line dimensions, each containing 2 elements “CHANGE”, “DONT CHANGE” and affecting the powerline status (connected / disconnected)

  • “set_bus”: dim_topo dimensions, each containing 4 choices: “DISCONNECT”, “DONT AFFECT”, “CONNECT TO BUSBAR 1”, or “CONNECT TO BUSBAR 2”, “CONNECT TO BUSBAR 3”, … and affecting to which busbar an object is connected

  • “change_bus”: dim_topo dimensions, each containing 2 choices: “CHANGE”, “DONT CHANGE” and affect to which busbar an element is connected

  • “redispatch”: sum(env.gen_redispatchable) dimensions, each containing a certain number of choices depending on the value of the keyword argument nb_bins[“redispatch”] (by default 7).

  • “curtail”: sum(env.gen_renewable) dimensions, each containing a certain number of choices depending on the value of the keyword argument nb_bins[“curtail”] (by default 7). This is the “conversion to discrete action” of the curtailment action.

  • “curtail_mw”: sum(env.gen_renewable) dimensions, completely equivalent to “curtail” for this representation. This is the “conversion to discrete action” of the curtailment action.

  • “set_storage”: n_storage dimensions, each containing a certain number of choices depending on the value of the keyword argument nb_bins[“set_storage”] (by default 7). This is the “conversion to discrete action” of the action on storage units.

  • “raise_alarm”: TODO

  • “raise_alert”: TODO

We offer some extra customization, with the keywords:

  • “sub_set_bus”: n_sub dimension. This type of representation encodes each different possible combination of elements that are possible at each substation. The choice at each component depends on the element connected at this substation. Only configurations that will not lead to straight game over will be generated.

  • “sub_change_bus”: n_sub dimension. Same comment as for “sub_set_bus”

  • “one_sub_set”: 1 single dimension. This type of representation differs from the previous one only by the fact that each step you can perform only one single action on a single substation (so unlikely to be illegal).

  • “one_sub_change”: 1 single dimension. Same as above.

  • “one_line_set”: 1 single dimension. In this type of representation, you have one dimension with 1 + 2 * n_line elements: first is “do nothing”, then next elements control the force connection or disconnection of the powerlines (new in version 1.10.0)

  • “one_line_change”: 1 single dimension. In this type of representation, you have 1 + n_line possibility for this element. First one is “do nothing” then it controls the change of status of any given line (new in version 1.10.0).

Warning

We recommend to use either “set” or “change” way to look at things (ie either you want to target a given state -in that case use “sub_set_bus”, “line_set_status”, “one_sub_set”, or “set_bus” __**OR**__ you prefer reasoning in terms of “i want to change this or that” in that case use “sub_change_bus”, “line_change_status”, “one_sub_change” or “change_bus”.

Combining a “set” and “change” on the same element will most likely lead to an “ambiguous action”. Indeed what grid2op can do if you “tell element A to go to bus 1” and “tell the same element A to switch to bus 2 if it was to 1 and to move to bus 1 if it was on bus 2”. It’s not clear at all (hence the “ambiguous”).

No error will be thrown if you mix this, this is your absolute right, be aware it might not lead to the result you expect though.

Note

The arguments “set_bus”, “sub_set_bus” and “one_sub_set” will all perform “set_bus” actions. The only difference if “how you represent these actions”:

  • In “set_bus” each component represent a single element of the grid. When you sample an action with this keyword you will possibly change all the elements of the grid at once (this is likely to be illega). Nothing prevents you to perform “weird” stuff, for example disconnecting a load or a generator (which is straight game over) or having a load or a generator that will be “alone” on a busbar (which will also lead to a straight game over). You can do anything with it, but as always “A great power comes with a great responsibility”.

  • In “sub_set_bus” each component represent a substation of the grid. When you sample an action from this, you will possibly change all the elements of the grid at once (because you can act on all the substation at the same time). As opposed to “set_bus” however this constraint the action space to “action that will not lead directly to a game over”, in practice.

  • In “one_sub_set”: the single component represent the whole grid. When you sample an action with this, you will sample a single action acting on a single substation. You will not be able to act on multiple substation with this.

For this reason, we also do not recommend using only one of these arguments and only provide only one of “set_bus”, “sub_set_bus” and “one_sub_set”. Again, no error will be thrown if you mix them but be warned that the resulting behaviour might not be what you expect.

Warning

The same as above holds for “change_bus”, “sub_change_bus” and “one_sub_change”: Use only one of these !

Danger

The keys set_bus and change_bus does not have the same meaning between this representation of the action and the DiscreteActSpace.

Warning

Depending on the presence absence of gymnasium and gym packages this class might behave differently.

In grid2op we tried to maintain compatibility both with gymnasium (newest) and gym (legacy, no more maintained) RL packages. The behaviour is the following:

See Gymnasium vs Gym for more information

Examples

If you simply want to use it you can do:

import grid2op
env_name = "l2rpn_case14_sandbox"  # or any other name
env = grid2op.make(env_name)

from grid2op.gym_compat import GymEnv, MultiDiscreteActSpace
gym_env = GymEnv(env)

gym_env.action_space = MultiDiscreteActSpace(env.action_space)

You can select the attribute you want to keep, for example:

gym_env.action_space = MultiDiscreteActSpace(env.observation_space,
                                             attr_to_keep=['redispatch', "curtail", "sub_set_bus"])

You can also apply some basic transformation when you “discretize” continuous action

gym_env.action_space = MultiDiscreteActSpace(env.observation_space,
                                             attr_to_keep=['redispatch', "curtail", "sub_set_bus"],
                                             nb_bins={"redispatch": 3, "curtail": 17},
                                             )

By default it is “discretized” in 7 different “bins”. The more “bins” there will be, the more “precise” you can be in your control, but the higher the dimension of the action space.

class grid2op.gym_compat.MultiDiscreteActSpaceLegacyGym(grid2op_action_space: ActionSpace, attr_to_keep: Tuple[Literal['set_line_status'], Literal['change_line_status'], Literal['set_bus'], Literal['sub_set_bus'], Literal['one_sub_set'], Literal['change_bus'], Literal['sub_change_bus'], Literal['one_sub_change'], Literal['redispatch'], Literal['set_storage'], Literal['curtail'], Literal['curtail_mw'], Literal['one_line_set'], Literal['one_line_change']] | None = ('set_line_status', 'change_line_status', 'set_bus', 'change_bus', 'redispatch', 'set_storage', 'curtail', 'raise_alarm', 'raise_alert'), nb_bins: Dict[Literal['redispatch', 'set_storage', 'curtail', 'curtail_mw'], int] | None = None)

This class allows to convert a grid2op action space into a gym “MultiDiscrete”. This means that the action are labeled, and instead of describing the action itself, you provide only its ID.

Note

This action space is particularly suited for represented discrete actions.

It is possible to represent continuous actions with it. In that case, the continuous actions are “binarized” thanks to the ContinuousToDiscreteConverter. Feel free to consult its documentation for more information.

In this case it will extract all the features in all the action with:

  • “set_line_status”: n_line dimensions, each containing 3 choices “DISCONNECT”, “DONT AFFECT”, “FORCE CONNECTION” and affecting the powerline status (connected / disconnected)

  • “change_line_status”: n_line dimensions, each containing 2 elements “CHANGE”, “DONT CHANGE” and affecting the powerline status (connected / disconnected)

  • “set_bus”: dim_topo dimensions, each containing 4 choices: “DISCONNECT”, “DONT AFFECT”, “CONNECT TO BUSBAR 1”, or “CONNECT TO BUSBAR 2”, “CONNECT TO BUSBAR 3”, … and affecting to which busbar an object is connected

  • “change_bus”: dim_topo dimensions, each containing 2 choices: “CHANGE”, “DONT CHANGE” and affect to which busbar an element is connected

  • “redispatch”: sum(env.gen_redispatchable) dimensions, each containing a certain number of choices depending on the value of the keyword argument nb_bins[“redispatch”] (by default 7).

  • “curtail”: sum(env.gen_renewable) dimensions, each containing a certain number of choices depending on the value of the keyword argument nb_bins[“curtail”] (by default 7). This is the “conversion to discrete action” of the curtailment action.

  • “curtail_mw”: sum(env.gen_renewable) dimensions, completely equivalent to “curtail” for this representation. This is the “conversion to discrete action” of the curtailment action.

  • “set_storage”: n_storage dimensions, each containing a certain number of choices depending on the value of the keyword argument nb_bins[“set_storage”] (by default 7). This is the “conversion to discrete action” of the action on storage units.

  • “raise_alarm”: TODO

  • “raise_alert”: TODO

We offer some extra customization, with the keywords:

  • “sub_set_bus”: n_sub dimension. This type of representation encodes each different possible combination of elements that are possible at each substation. The choice at each component depends on the element connected at this substation. Only configurations that will not lead to straight game over will be generated.

  • “sub_change_bus”: n_sub dimension. Same comment as for “sub_set_bus”

  • “one_sub_set”: 1 single dimension. This type of representation differs from the previous one only by the fact that each step you can perform only one single action on a single substation (so unlikely to be illegal).

  • “one_sub_change”: 1 single dimension. Same as above.

  • “one_line_set”: 1 single dimension. In this type of representation, you have one dimension with 1 + 2 * n_line elements: first is “do nothing”, then next elements control the force connection or disconnection of the powerlines (new in version 1.10.0)

  • “one_line_change”: 1 single dimension. In this type of representation, you have 1 + n_line possibility for this element. First one is “do nothing” then it controls the change of status of any given line (new in version 1.10.0).

Warning

We recommend to use either “set” or “change” way to look at things (ie either you want to target a given state -in that case use “sub_set_bus”, “line_set_status”, “one_sub_set”, or “set_bus” __**OR**__ you prefer reasoning in terms of “i want to change this or that” in that case use “sub_change_bus”, “line_change_status”, “one_sub_change” or “change_bus”.

Combining a “set” and “change” on the same element will most likely lead to an “ambiguous action”. Indeed what grid2op can do if you “tell element A to go to bus 1” and “tell the same element A to switch to bus 2 if it was to 1 and to move to bus 1 if it was on bus 2”. It’s not clear at all (hence the “ambiguous”).

No error will be thrown if you mix this, this is your absolute right, be aware it might not lead to the result you expect though.

Note

The arguments “set_bus”, “sub_set_bus” and “one_sub_set” will all perform “set_bus” actions. The only difference if “how you represent these actions”:

  • In “set_bus” each component represent a single element of the grid. When you sample an action with this keyword you will possibly change all the elements of the grid at once (this is likely to be illega). Nothing prevents you to perform “weird” stuff, for example disconnecting a load or a generator (which is straight game over) or having a load or a generator that will be “alone” on a busbar (which will also lead to a straight game over). You can do anything with it, but as always “A great power comes with a great responsibility”.

  • In “sub_set_bus” each component represent a substation of the grid. When you sample an action from this, you will possibly change all the elements of the grid at once (because you can act on all the substation at the same time). As opposed to “set_bus” however this constraint the action space to “action that will not lead directly to a game over”, in practice.

  • In “one_sub_set”: the single component represent the whole grid. When you sample an action with this, you will sample a single action acting on a single substation. You will not be able to act on multiple substation with this.

For this reason, we also do not recommend using only one of these arguments and only provide only one of “set_bus”, “sub_set_bus” and “one_sub_set”. Again, no error will be thrown if you mix them but be warned that the resulting behaviour might not be what you expect.

Warning

The same as above holds for “change_bus”, “sub_change_bus” and “one_sub_change”: Use only one of these !

Danger

The keys set_bus and change_bus does not have the same meaning between this representation of the action and the DiscreteActSpace.

Warning

Depending on the presence absence of gymnasium and gym packages this class might behave differently.

In grid2op we tried to maintain compatibility both with gymnasium (newest) and gym (legacy, no more maintained) RL packages. The behaviour is the following:

See Gymnasium vs Gym for more information

Examples

If you simply want to use it you can do:

import grid2op
env_name = "l2rpn_case14_sandbox"  # or any other name
env = grid2op.make(env_name)

from grid2op.gym_compat import GymEnv, MultiDiscreteActSpace
gym_env = GymEnv(env)

gym_env.action_space = MultiDiscreteActSpace(env.action_space)

You can select the attribute you want to keep, for example:

gym_env.action_space = MultiDiscreteActSpace(env.observation_space,
                                             attr_to_keep=['redispatch', "curtail", "sub_set_bus"])

You can also apply some basic transformation when you “discretize” continuous action

gym_env.action_space = MultiDiscreteActSpace(env.observation_space,
                                             attr_to_keep=['redispatch', "curtail", "sub_set_bus"],
                                             nb_bins={"redispatch": 3, "curtail": 17},
                                             )

By default it is “discretized” in 7 different “bins”. The more “bins” there will be, the more “precise” you can be in your control, but the higher the dimension of the action space.

grid2op.gym_compat.MultiToTupleConverter

alias of MultiToTupleConverterGymnasium

class grid2op.gym_compat.MultiToTupleConverterGymnasium(init_space=None)

Some framework, for example ray[rllib] do not support MultiBinary nor MultiDiscrete gym action space. Apparently this is not going to change in a near future (see https://github.com/ray-project/ray/issues/1519).

We choose to encode some variable using MultiBinary variable in grid2op. This allows for easy manipulation of them if using these frameworks.

MultiBinary are encoded with gym Tuple of gym Discrete variables.

Warning

Depending on the presence absence of gymnasium and gym packages this class might behave differently.

In grid2op we tried to maintain compatibility both with gymnasium (newest) and gym (legacy, no more maintained) RL packages. The behaviour is the following:

See Gymnasium vs Gym for more information

TODO add code example

class grid2op.gym_compat.MultiToTupleConverterLegacyGym(init_space=None)

Some framework, for example ray[rllib] do not support MultiBinary nor MultiDiscrete gym action space. Apparently this is not going to change in a near future (see https://github.com/ray-project/ray/issues/1519).

We choose to encode some variable using MultiBinary variable in grid2op. This allows for easy manipulation of them if using these frameworks.

MultiBinary are encoded with gym Tuple of gym Discrete variables.

Warning

Depending on the presence absence of gymnasium and gym packages this class might behave differently.

In grid2op we tried to maintain compatibility both with gymnasium (newest) and gym (legacy, no more maintained) RL packages. The behaviour is the following:

See Gymnasium vs Gym for more information

TODO add code example

grid2op.gym_compat.ScalerAttrConverter

alias of ScalerAttrConverterGymnasium

class grid2op.gym_compat.ScalerAttrConverterGymnasium(substract, divide, dtype=None, init_space=None)

This is a scaler that transforms a initial gym space init_space into its scale version.

It can be use to scale the observation by substracting the mean and dividing by the variance for example.

TODO work in progress !

Need help if you can :-)

Warning

Depending on the presence absence of gymnasium and gym packages this class might behave differently.

In grid2op we tried to maintain compatibility both with gymnasium (newest) and gym (legacy, no more maintained) RL packages. The behaviour is the following:

See Gymnasium vs Gym for more information

class grid2op.gym_compat.ScalerAttrConverterLegacyGym(substract, divide, dtype=None, init_space=None)

This is a scaler that transforms a initial gym space init_space into its scale version.

It can be use to scale the observation by substracting the mean and dividing by the variance for example.

TODO work in progress !

Need help if you can :-)

Warning

Depending on the presence absence of gymnasium and gym packages this class might behave differently.

In grid2op we tried to maintain compatibility both with gymnasium (newest) and gym (legacy, no more maintained) RL packages. The behaviour is the following:

See Gymnasium vs Gym for more information

grid2op.gym_compat.gym_space_converter._BaseGymSpaceConverter

alias of _BaseGymnasiumSpaceConverter

If you still can’t find what you’re looking for, try in one of the following pages:

Still trouble finding the information ? Do not hesitate to send a github issue about the documentation at this link: Documentation issue template