def __init__(
self,
envs_dir,
_add_to_name="", # internal, for test only, do not use !
**kwargs):
GridObjects.__init__(self)
RandomObject.__init__(self)
self.current_env = None
self.env_index = None
self.mix_envs = []
# Special case handling for backend
backendClass = None
if "backend" in kwargs:
backendClass = type(kwargs["backend"])
del kwargs["backend"]
# Inline import to prevent cyclical import
from grid2op.MakeEnv.Make import make
try:
for env_dir in sorted(os.listdir(envs_dir)):
env_path = os.path.join(envs_dir, env_dir)
if not os.path.isdir(env_path):
continue
# Special case for backend
if backendClass is not None:
env = make(env_path,
backend=backendClass(),
_add_to_name=_add_to_name,
**kwargs)
else:
env = make(env_path, **kwargs)
self.mix_envs.append(env)
except Exception as e:
err_msg = "MultiMix environment creation failed: {}".format(e)
raise EnvError(err_msg)
if len(self.mix_envs) == 0:
err_msg = "MultiMix envs_dir did not contain any valid env"
raise EnvError(err_msg)
self.env_index = 0
self.current_env = self.mix_envs[self.env_index]
# Make sure GridObject class attributes are set from first env
# Should be fine since the grid is the same for all envs
multi_env_name = os.path.basename(
os.path.abspath(envs_dir)) + _add_to_name
save_env_name = self.current_env.env_name
self.current_env.env_name = multi_env_name
self.__class__ = self.init_grid(self.current_env)
self.current_env.env_name = save_env_name
def main():
# Initialize the environment and agent
path_grid = "rte_case14_redisp"
env = make(path_grid,
reward_class=L2RPNReward,
action_class=TopologyChangeAction)
# my_agent = DoNothingAgent(env.action_space) # Acts as the baseline agent
my_agent = DeepQAgent(env.action_space)
num_states = my_agent.convert_obs(env.reset()).shape[0]
num_actions = my_agent.action_space.size()
num_training_iterations = 5000
num_run_iterations = 5000
print('State space size:', num_states)
print('Action space size:', num_actions)
print('Training iterations:', num_training_iterations)
print('Run iterations:', num_run_iterations)
# Plot grid visualization
plot_grid_layout(env)
# # Load an existing network
# my_agent.id = '{}_{}_{}_il'.format(path_grid, my_agent.__class__.__name__, num_training_iterations)
# my_agent.id = '{}_{}_{}_test'.format(path_grid, my_agent.__class__.__name__, num_training_iterations)
# my_agent.init_deep_q(my_agent.convert_obs(env.reset()))
# my_agent.load(os.path.join('saved_networks', my_agent.id))
# # Load Imitation Learning network
# num_samples = 1000
# run_id = 0
# il_network_path = '{}_{}_{}_il'.format(path_grid, num_samples, run_id)
# my_agent.init_deep_q(my_agent.convert_obs(env.reset()))
# my_agent.load(os.path.join('saved_networks', 'imitation_learning', il_network_path))
# Train a new agent
my_agent.id = '{}_{}_{}'.format(path_grid, my_agent.__class__.__name__,
num_training_iterations)
train_agent(my_agent, env, num_iterations=num_training_iterations)
# Evaluate the agent
path_grid = "rte_case14_realistic"
env = make(path_grid,
reward_class=L2RPNReward,
action_class=TopologyChangeAction)
run_agent(env,
my_agent,
num_iterations=num_run_iterations,
plot_replay_episodes=True,
use_runner=False)
""" Convert a CompleteObservation object from Grid2Op to a vector as input for the Q-network.
This method equals the convert_obs method in DeepQAgent. """
return np.concatenate((
observation.prod_p / 150,
observation.load_p / 120,
observation.rho / 2,
observation.timestep_overflow / 10,
observation.line_status,
(observation.topo_vect + 1) / 3,
observation.time_before_cooldown_line / 10,
observation.time_before_cooldown_sub / 10))
# Setup the environment
path_grid = "rte_case14_redisp"
env = make(path_grid, reward_class=L2RPNReward, action_class=TopologyChangeAction)
obs = env.reset()
run_id = 0
n = 1000
num_states = convert_obs(obs).shape[0]
num_actions = 191 # Specific for TopologyChangeAction on case 14
print('State space size:', num_states)
print('Action space size:', num_actions)
converter = IdToAct(env.action_space)
converter.init_converter()
states = np.zeros((n, num_states))
rewards = np.zeros((n, num_actions))
cum_reward = 0.
reset_count = 0
def __init__(
self,
envs_dir,
experimental_read_from_local_dir=False,
_add_to_name="", # internal, for test only, do not use !
_compat_glop_version=None, # internal, for test only, do not use !
_test=False,
**kwargs):
GridObjects.__init__(self)
RandomObject.__init__(self)
self.current_env = None
self.env_index = None
self.mix_envs = []
self._env_dir = os.path.abspath(envs_dir)
# Special case handling for backend
# TODO: with backend.copy() instead !
backendClass = None
if "backend" in kwargs:
backendClass = type(kwargs["backend"])
del kwargs["backend"]
# Inline import to prevent cyclical import
from grid2op.MakeEnv.Make import make
# TODO reuse same observation_space and action_space in all the envs maybe ?
try:
for env_dir in sorted(os.listdir(envs_dir)):
env_path = os.path.join(envs_dir, env_dir)
if not os.path.isdir(env_path):
continue
# Special case for backend
if backendClass is not None:
env = make(env_path,
backend=backendClass(),
_add_to_name=_add_to_name,
_compat_glop_version=_compat_glop_version,
test=_test,
experimental_read_from_local_dir=
experimental_read_from_local_dir,
**kwargs)
else:
env = make(env_path,
_add_to_name=_add_to_name,
_compat_glop_version=_compat_glop_version,
test=_test,
experimental_read_from_local_dir=
experimental_read_from_local_dir,
**kwargs)
self.mix_envs.append(env)
except Exception as exc_:
err_msg = "MultiMix environment creation failed: {}".format(exc_)
raise EnvError(err_msg)
if len(self.mix_envs) == 0:
err_msg = "MultiMix envs_dir did not contain any valid env"
raise EnvError(err_msg)
self.env_index = 0
self.current_env = self.mix_envs[self.env_index]
# Make sure GridObject class attributes are set from first env
# Should be fine since the grid is the same for all envs
multi_env_name = os.path.basename(
os.path.abspath(envs_dir)) + _add_to_name
save_env_name = self.current_env.env_name
self.current_env.env_name = multi_env_name
self.__class__ = self.init_grid(self.current_env)
self.current_env.env_name = save_env_name
