Source code for grid2op.Observation.observationSpace

# Copyright (c) 2019-2020, RTE (
# See AUTHORS.txt
# This Source Code Form is subject to the terms of the Mozilla Public License, version 2.0.
# If a copy of the Mozilla Public License, version 2.0 was not distributed with this file,
# you can obtain one at
# SPDX-License-Identifier: MPL-2.0
# This file is part of Grid2Op, Grid2Op a testbed platform to model sequential decision making in power systems.

import sys
import copy
import logging
from grid2op.Exceptions.EnvExceptions import EnvError

from grid2op.Observation.serializableObservationSpace import (
from grid2op.Reward import RewardHelper
from grid2op.Observation.completeObservation import CompleteObservation

[docs]class ObservationSpace(SerializableObservationSpace): """ Helper that provides useful functions to manipulate :class:`BaseObservation`. BaseObservation should only be built using this Helper. It is absolutely not recommended to make an observation directly form its constructor. This class represents the same concept as the "BaseObservation Space" in the OpenAI gym framework. Attributes ---------- with_forecast: ``bool`` If ``True`` the :func:`BaseObservation.simulate` will be available. If ``False`` it will deactivate this possibility. If `simulate` function is not used, setting it to ``False`` can lead to non neglectible speed-ups. observationClass: ``type`` Class used to build the observations. It defaults to :class:`CompleteObservation` _simulate_parameters: :class:`grid2op.Parameters.Parameters` Type of Parameters used to compute powerflow for the forecast. rewardClass: ``type`` Class used by the :class:`grid2op.Environment.Environment` to send information about its state to the :class:`grid2op.BaseAgent.BaseAgent`. You can change this class to differentiate between the reward of output of :func:`BaseObservation.simulate` and the reward used to train the BaseAgent. action_helper_env: :class:`grid2op.Action.ActionSpace` BaseAction space used to create action during the :func:`BaseObservation.simulate` reward_helper: :class:`grid2op.Reward.HelperReward` BaseReward function used by the the :func:`BaseObservation.simulate` function. obs_env: :class:`_ObsEnv` Instance of the environment used by the BaseObservation Helper to provide forcecast of the grid state. _empty_obs: :class:`BaseObservation` An instance of the observation with appropriate dimensions. It is updated and will be sent to he BaseAgent. """
[docs] def __init__( self, gridobj, env, rewardClass=None, observationClass=CompleteObservation, actionClass=None, with_forecast=True, kwargs_observation=None, logger=None, ): """ INTERNAL .. warning:: /!\\\\ Internal, do not use unless you know what you are doing /!\\\\ Env: requires :attr:`grid2op.Environment.parameters` and :attr:`grid2op.Environment.backend` to be valid """ from grid2op.Environment._ObsEnv import ( _ObsEnv, ) # lazy import to prevent circular references (Env -> Observation -> Obs Space -> _ObsEnv -> Env) if actionClass is None: from grid2op.Action import CompleteAction actionClass = CompleteAction if logger is None: self.logger = logging.getLogger(__name__) self.logger.disabled = True else: self.logger: logging.Logger = logger.getChild("grid2op_ObsSpace") SerializableObservationSpace.__init__( self, gridobj, observationClass=observationClass ) self.with_forecast = with_forecast self._simulate_parameters = copy.deepcopy(env.parameters) self._legal_action = env._game_rules.legal_action self._env_param = copy.deepcopy(env.parameters) if rewardClass is None: self._reward_func = env._reward_helper.template_reward else: self._reward_func = rewardClass # helpers self.action_helper_env = env._helper_action_env self.reward_helper = RewardHelper(reward_func=self._reward_func, logger=self.logger) self.reward_helper.initialize(env) other_rewards = {k: v.rewardClass for k, v in env.other_rewards.items()} # TODO here: have another backend class maybe if env.backend._can_be_copied: try: self._backend_obs = env.backend.copy() except Exception as exc_: self._backend_obs = None self.logger.warn(f"Backend cannot be copied, simulate feature will " f"be unsusable. Error was: {exc_}") self._deactivate_simulate(env) else: self._backend_obs = None self._deactivate_simulate(env) _ObsEnv_class = _ObsEnv.init_grid( type(env.backend), force_module=_ObsEnv.__module__ ) _ObsEnv_class._INIT_GRID_CLS = _ObsEnv # otherwise it's lost setattr(sys.modules[_ObsEnv.__module__], _ObsEnv_class.__name__, _ObsEnv_class) self.obs_env = _ObsEnv_class( init_env_path=None, # don't leak the path of the real grid to the observation space init_grid_path=None, # don't leak the path of the real grid to the observation space backend_instanciated=self._backend_obs, obsClass=CompleteObservation, # do not put self.observationClass otherwise it's initialized twice parameters=self._simulate_parameters, reward_helper=self.reward_helper, action_helper=self.action_helper_env, thermal_limit_a=env.get_thermal_limit(), legalActClass=copy.deepcopy(env._legalActClass), other_rewards=other_rewards, helper_action_class=env._helper_action_class, helper_action_env=env._helper_action_env, epsilon_poly=env._epsilon_poly, tol_poly=env._tol_poly, has_attention_budget=env._has_attention_budget, attention_budget_cls=env._attention_budget_cls, kwargs_attention_budget=env._kwargs_attention_budget, max_episode_duration=env.max_episode_duration(), logger=self.logger, _complete_action_cls=env._complete_action_cls, _ptr_orig_obs_space=self, ) for k, v in self.obs_env.other_rewards.items(): v.initialize(env) self._empty_obs = self._template_obj self._update_env_time = 0.0 self.__nb_simulate_called_this_step = 0 self.__nb_simulate_called_this_episode = 0 # extra argument to build the observation if kwargs_observation is None: kwargs_observation = {} self._ptr_kwargs_observation = kwargs_observation
def _deactivate_simulate(self, env): self._backend_obs = None self.with_forecast = False env.deactivate_forecast() env.backend._can_be_copied = False self.logger.warn("Forecasts have been deactivated because " "the backend cannot be copied.")
[docs] def simulate_called(self): """ INTERNAL .. warning:: /!\\\\ Internal, do not use unless you know what you are doing /!\\\\ Tells this class that the "obs.simulate" function has been called. """ self.__nb_simulate_called_this_step += 1 self.__nb_simulate_called_this_episode += 1
@property def nb_simulate_called_this_episode(self): return self.__nb_simulate_called_this_episode @property def nb_simulate_called_this_step(self): return self.__nb_simulate_called_this_step
[docs] def can_use_simulate(self) -> bool: """ This checks on the rules if the agent has not made too many calls to "obs.simulate" this step """ return self._legal_action.can_use_simulate( self.__nb_simulate_called_this_step, self.__nb_simulate_called_this_episode, self._env_param, )
def _change_parameters(self, new_param): """ INTERNAL .. warning:: /!\\\\ Internal, do not use unless you know what you are doing /!\\\\ change the parameter of the "simulate" environment """ self.obs_env.change_parameters(new_param) self._simulate_parameters = new_param
[docs] def change_other_rewards(self, dict_reward): """ this function is used to change the "other rewards" used when you perform simulate. This can be used, for example, when you want to do faster call to "simulate". In this case you can remove all the "other_rewards" that will be used by the simulate function. Parameters ---------- dict_reward: ``dict`` see description of :attr:`grid2op.Environment.BaseEnv.other_rewards` Examples --------- If you want to deactivate the reward in the simulate function, you can do as following: .. code-block:: python import grid2op from grid2op.Reward import CloseToOverflowReward, L2RPNReward, RedispReward env_name = "l2rpn_case14_sandbox" other_rewards = {"close_overflow": CloseToOverflowReward, "l2rpn": L2RPNReward, "redisp": RedispReward} env = grid2op.make(env_name, other_rewards=other_rewards) env.observation_space.change_other_rewards({}) """ from grid2op.Reward import BaseReward from grid2op.Exceptions import Grid2OpException self.obs_env.other_rewards = {} for k, v in dict_reward.items(): if not issubclass(v, BaseReward): raise Grid2OpException( 'All values of "rewards" key word argument should be classes that inherit ' 'from "grid2op.BaseReward"' ) if not isinstance(k, str): raise Grid2OpException( 'All keys of "rewards" should be of string type.' ) self.obs_env.other_rewards[k] = RewardHelper(v) for k, v in self.obs_env.other_rewards.items(): v.initialize(self.obs_env)
def change_reward(self, reward_func): if self.obs_env.is_valid(): self.obs_env._reward_helper.change_reward(reward_func) else: raise EnvError("Impossible to change the reward of the simulate " "function when you cannot simulate (because the " "backend could not be copied)") def reset_space(self): if self.with_forecast: if self.obs_env.is_valid(): self.obs_env.reset_space() else: raise EnvError("Impossible to reset_space " "function when you cannot simulate (because the " "backend could not be copied)") self.action_helper_env.actionClass.reset_space()
[docs] def __call__(self, env, _update_state=True): if self.with_forecast: self.obs_env.update_grid(env) res = self.observationClass( obs_env=self.obs_env if self.obs_env.is_valid() else None, action_helper=self.action_helper_env, random_prng=self.space_prng, **self._ptr_kwargs_observation ) self.__nb_simulate_called_this_step = 0 if _update_state: # TODO how to make sure that whatever the number of time i call "simulate" i still get the same observations # TODO use self.obs_prng when updating actions res.update(env=env, with_forecast=self.with_forecast) return res
[docs] def size_obs(self): """ Size if the observation vector would be flatten :return: """ return self.n
[docs] def get_empty_observation(self): """ INTERNAL .. warning:: /!\\\\ Internal, do not use unless you know what you are doing /!\\\\ return an empty observation, for internal use only. """ return copy.deepcopy(self._empty_obs)
[docs] def reset(self, real_env): """reset the observation space with the new values of the environment""" self.obs_env._reward_helper.reset(real_env) self.__nb_simulate_called_this_step = 0 self.__nb_simulate_called_this_episode = 0 for k, v in self.obs_env.other_rewards.items(): v.reset(real_env) self._env_param = copy.deepcopy(real_env.parameters)
def _custom_deepcopy_for_copy(self, new_obj): """implements a faster "res = copy.deepcopy(self)" to use in "self.copy" Do not use it anywhere else... """ # TODO clean that after it is working... (ie make this method per class...) # fill the super classes super()._custom_deepcopy_for_copy(new_obj) # now fill my class new_obj.with_forecast = self.with_forecast new_obj._simulate_parameters = copy.deepcopy(self._simulate_parameters) new_obj._reward_func = copy.deepcopy(self._reward_func) new_obj.action_helper_env = self.action_helper_env # const new_obj.reward_helper = copy.deepcopy(self.reward_helper) new_obj._backend_obs = self._backend_obs # ptr to a backend for simulate new_obj.obs_env = self.obs_env # it is None anyway ! new_obj._update_env_time = self._update_env_time new_obj.__nb_simulate_called_this_step = self.__nb_simulate_called_this_step new_obj.__nb_simulate_called_this_episode = ( self.__nb_simulate_called_this_episode ) new_obj._env_param = copy.deepcopy(self._env_param) # as it's a "pointer" it's init from the env when needed here # this is why i don't deep copy it here ! new_obj._ptr_kwargs_observation = self._ptr_kwargs_observation
[docs] def copy(self, copy_backend=False): """ INTERNAL .. warning:: /!\\\\ Internal, do not use unless you know what you are doing /!\\\\ Perform a deep copy of the Observation space. """ backend = self._backend_obs self._backend_obs = None obs_ = self._empty_obs self._empty_obs = None obs_env = self.obs_env self.obs_env = None # performs the copy # res = copy.deepcopy(self) # painfully slow... # create an empty "me" my_cls = type(self) res = my_cls.__new__(my_cls) self._custom_deepcopy_for_copy(res) if not copy_backend: res._backend_obs = backend res._empty_obs = obs_.copy() res.obs_env = obs_env else: res.obs_env = obs_env.copy() res.obs_env._ptr_orig_obs_space = res res._backend_obs = res.obs_env.backend res._empty_obs = obs_.copy() res._empty_obs._obs_env = res.obs_env # assign back the results self._backend_obs = backend self._empty_obs = obs_ self.obs_env = obs_env return res
def close(self): if self.obs_env is not None: self.obs_env.close() del self.obs_env self.obs_env = None