Skip to content

Latest commit

 

History

History
563 lines (422 loc) · 26.3 KB

gym.rst

File metadata and controls

563 lines (422 loc) · 26.3 KB
Raw

grid2op.gym_compat

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Starting with 1.9.1 we switch (as advised) from the legacy "gym" framework to the new "gymnasium" framework (gym is no longer maintained since v0.26.2, see https://www.gymlibrary.dev/).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_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 gym_compat_box_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:

Table of Contents

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_module 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 gym_compat_box_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 base_gym_space_function

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

Recommended usage of grid2op with other framework

Reinforcement learning frameworks

TODO

Any contribution is welcome here

Other frameworks

Any contribution is welcome here too (-:

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

grid2op.gym_compat

grid2op.gym_compat.gym_space_converter._BaseGymSpaceConverter