It is recommended to have a look at the 0_basic_functionalities, 1_Observation_Agents and 2_Action_GridManipulation notebooks before getting into this one.
Objectives
In this notebook we will expose :
NB for this tutorial we train an Agent inspired from this blog post: deep-reinforcement-learning-tutorial-with-open-ai-gym. Many other different reinforcement learning tutorial exist. The code showed in this notebook has no pretention except to demonstrate how to use Grid2Op functionality to train a Deep Reinforcement learning Agent and inspect its behaviour. There are absolutely nothing implied about the performance, training strategy, type of Agent, meta parameters etc. All of them are purely "random".
import os
import sys
import grid2op
res = None
try:
from jyquickhelper import add_notebook_menu
res = add_notebook_menu()
except ModuleNotFoundError:
print("Impossible to automatically add a menu / table of content to this notebook.\nYou can download \"jyquickhelper\" package with: \n\"pip install jyquickhelper\"")
res
Grid2op package has been built with an "object oriented" perspective: almost everything is encapsulated in a dedicated class
. This allows for more customization of the plateform.
The downside of this approach is that machine learning method, and especially deep learning, often prefers to deal with vectors rather than with complex
objects. Indeed, as we covered in the previous tutorials on the platform, building our own actions can be tedious and can sometime require knowledge of the powergrid.
On the contrary, in most of standard Reinforcement Learning environment, actions have an higher representation. For example in pacman, there are 4 different types of actions: turn left, turn right, go up or do down. This allows for easy sampling (you need to achieve a uniform sampling you simply need to sample a number between 0 and 3 included) and an easy representation: each action is a different component of a vector of dimension 4 [because there are 4 actions].
On the other hand this representation is not "human friendly". It is quite convenient in the case of pacman because the action space is rather small making it possible to remember which action corresponds to which component, but in the case of the grid2op package, there are hundreds, sometimes thousands of actions, making it impossible to remember which component corresponds to which actions. We suppose we don't really care about this fact here, as tutorials on Reinforcement Learning with discrete action space often assume that actions are labelled with integer (such as in pacman for example).
Howerever, to allow the training of RL agent more easily, we allows to make some "Converters" whose roles are to allow an agent to deal with a custom representation of the action space. The class AgentWithConverter is perfect for such usage.
# import the usefull class
import numpy as np
from grid2op import make_new
from grid2op.Agent import RandomAgent
max_iter = 100 # to make computation much faster we will only consider 50 time steps instead of 287
env = make_new("rte_case14_redisp", test=True)
env.seed(0) # this is to ensure the same action are taken by the "RandomAgent".
my_agent = RandomAgent(env.action_space)
/home/donnotben/Documents/Grid2Op_dev/getting_started/grid2op/MakeEnv/MakeNew.py:138: UserWarning: You are using a development environment. This is really not recommended for training agents.
And that's it. This agent will be able to perform any action, but instead of going through the description of the actions from a powersystem point of view (ie setting what is connected to what, what is disconnected etc.) it will simply choose an integer with the method my_act
this integer will then be converter back to a proper valid action.
Here we have an example on the action representation as seen by the Agent:
for el in range(3):
print(my_agent.my_act(None, None))
172 47 117
And below you can see the "act
" functions behaves as expected:
for el in range(3):
print(my_agent.act(None, None))
This action will: - NOT change anything to the injections - NOT perform any redispatching action - NOT force any line status - NOT switch any line status - NOT switch anything in the topology - Set the bus of the following element: - assign bus 2 to line (origin) 12 [on substation 5] - assign bus 2 to line (origin) 13 [on substation 5] - assign bus 1 to line (origin) 14 [on substation 5] - assign bus 2 to line (extremity) 17 [on substation 5] - assign bus 1 to generator 2 [on substation 5] - assign bus 1 to load 5 [on substation 5] This action will: - NOT change anything to the injections - NOT perform any redispatching action - NOT force any line status - NOT switch any line status - Change the bus of the following element: - switch bus of line (origin) 3 [on substation 8] - switch bus of line (extremity) 16 [on substation 8] - switch bus of line (extremity) 19 [on substation 8] - NOT force any particular bus configuration This action will: - NOT change anything to the injections - NOT perform any redispatching action - NOT force any line status - NOT switch any line status - Change the bus of the following element: - switch bus of generator 1 [on substation 2] - switch bus of load 1 [on substation 2] - NOT force any particular bus configuration
NB lots of these actions are equivalent to the "do nothing" action at some point. For example, when trying to reconnect a powerline that is already connected will do nothing. Same for the topology. If everything is already connected to bus 1, then the action to connect things to bus 1 on the same substation will not affect the powergrid.
For this tutorial, we will expose to built a Q-learning Agent. Most of the code (except the neural network architecture) are inspired from this blog post: https://towardsdatascience.com/deep-reinforcement-learning-tutorial-with-open-ai-gym-c0de4471f368.
Requirements This notebook require to have keras
installed on your machine.
As always in these notebook, we will use the case14_fromfile
Environment. No proper care has been taken to set the thermal limits on this grid. It's unlikely that the agent can learn anything in this context.
The type of Agent were are using require a bit of set up, independantly of Grid2Op. We will reuse the code showed in https://towardsdatascience.com/deep-reinforcement-learning-tutorial-with-open-ai-gym-c0de4471f368 and in Reinforcement-Learning-Tutorial from Abhinav Sagar code under a MIT license found here: MIT License.
This first section is here to define these classes.
But first let's import the necessary dependencies
#tf2.0 friendly
import numpy as np
import random
import warnings
with warnings.catch_warnings():
warnings.filterwarnings("ignore", category=FutureWarning)
import tensorflow.keras
import tensorflow.keras.backend as K
from tensorflow.keras.models import load_model, Sequential, Model
from tensorflow.keras.optimizers import Adam
from tensorflow.keras.layers import Activation, Dropout, Flatten, Dense, subtract, add
from tensorflow.keras.layers import Input, Lambda, Concatenate
First we define a "replay buffer" necessary to train the Agent.
# Credit Abhinav Sagar:
# https://github.com/abhinavsagar/Reinforcement-Learning-Tutorial
# Code under MIT license, available at:
# https://github.com/abhinavsagar/Reinforcement-Learning-Tutorial/blob/master/LICENSE
from collections import deque
class ReplayBuffer:
"""Constructs a buffer object that stores the past moves
and samples a set of subsamples"""
def __init__(self, buffer_size):
self.buffer_size = buffer_size
self.count = 0
self.buffer = deque()
def add(self, s, a, r, d, s2):
"""Add an experience to the buffer"""
# S represents current state, a is action,
# r is reward, d is whether it is the end,
# and s2 is next state
experience = (s, a, r, d, s2)
if self.count < self.buffer_size:
self.buffer.append(experience)
self.count += 1
else:
self.buffer.popleft()
self.buffer.append(experience)
def size(self):
return self.count
def sample(self, batch_size):
"""Samples a total of elements equal to batch_size from buffer
if buffer contains enough elements. Otherwise return all elements"""
batch = []
if self.count < batch_size:
batch = random.sample(self.buffer, self.count)
else:
batch = random.sample(self.buffer, batch_size)
# Maps each experience in batch in batches of states, actions, rewards
# and new states
s_batch, a_batch, r_batch, d_batch, s2_batch = list(map(np.array, list(zip(*batch))))
return s_batch, a_batch, r_batch, d_batch, s2_batch
def clear(self):
self.buffer.clear()
self.count = 0
Then we re-use the default parameters, note that these can be optimized. Nothing has been changed for this example.
For more information about them, please refer to the blog post of Abhinav Sagar available here.
DECAY_RATE = 0.9
BUFFER_SIZE = 40000
MINIBATCH_SIZE = 64
TOT_FRAME = 3000000
EPSILON_DECAY = 10000
MIN_OBSERVATION = 42 #5000
FINAL_EPSILON = 1/300 # have on average 1 random action per scenario of approx 287 time steps
INITIAL_EPSILON = 0.1
TAU = 0.01
ALPHA = 1
# Number of frames to "throw" into network
NUM_FRAMES = 1 ## this has been changed compared to the original implementation.
In the original code, the models were used to play an Atari game and the inputs were images. For our system, the inputs are "Observation" converted as vector.
For a more detailed description of the code used, please check:
This is why we adapted the original code from Abhinav Sagar:
First we extract relevant information about the dimension of the observation space, and the action space.
OBSERVATION_SIZE = env.observation_space.size()
NUM_ACTIONS = my_agent.action_space.n
A few comments here.
First, we don't change anything to the observation space. This means that the vector the agent will receive is really big, not scaled and with lots of informations that are not really usefull.
The code of the neural networks, used to solve this problem is modified very slightly to adapt. The biggest changes come from removing the convolutional layers, as well as adapting the input and output size.
For each of the method below, we specify what have been adapted/modified in comparison with the original blog post by Abhinav.
We want to emphasize here that these models are used through the "predict_movement
" method. This method outputs an integer: the action id that maximizes the expected reward at the current tim step in an episode of the game during the training process. It's perfectly suited to use a representation of actions with integer rather than with complete descriptions of what the agent is doing.
In the "reference" article https://towardsdatascience.com/deep-reinforcement-learning-tutorial-with-open-ai-gym-c0de4471f368, the author Abhinav Sagar made a dedicated environment based on SpaceInvader in the gym repository. We proceed here on a similar way, but with a the grid2op environment.
# Credit Abhinav Sagar:
# https://github.com/abhinavsagar/Reinforcement-Learning-Tutorial
# Code under MIT license, available at:
# https://github.com/abhinavsagar/Reinforcement-Learning-Tutorial/blob/master/LICENSE
class DeepQ(object):
"""Constructs the desired deep q learning network"""
def __init__(self, action_size, lr=1e-5, observation_size=OBSERVATION_SIZE):
# It is not modified from Abhinav Sagar's code, except for adding the possibility to change the learning rate
# in parameter is also present the size of the action space
# (it used to be a global variable in the original code)
self.action_size = action_size
self.observation_size = observation_size
self.model = None
self.target_model = None
self.lr_ = lr
self.qvalue_evolution = []
self.construct_q_network()
def construct_q_network(self):
# replacement of the Convolution layers by Dense layers, and change the size of the input space and output space
# Uses the network architecture found in DeepMind paper
self.model = Sequential()
input_layer = Input(shape = (self.observation_size*NUM_FRAMES,))
layer1 = Dense(self.observation_size*NUM_FRAMES)(input_layer)
layer1 = Activation('relu')(layer1)
layer2 = Dense(self.observation_size)(layer1)
layer2 = Activation('relu')(layer2)
layer3 = Dense(self.observation_size)(layer2)
layer3 = Activation('relu')(layer3)
layer4 = Dense(2*NUM_ACTIONS)(layer3)
layer4 = Activation('relu')(layer4)
output = Dense(NUM_ACTIONS)(layer4)
self.model = Model(inputs=[input_layer], outputs=[output])
self.model.compile(loss='mse', optimizer=Adam(lr=self.lr_))
self.target_model = Model(inputs=[input_layer], outputs=[output])
self.target_model.compile(loss='mse', optimizer=Adam(lr=self.lr_))
self.target_model.set_weights(self.model.get_weights())
def predict_movement(self, data, epsilon):
"""Predict movement of game controler where is epsilon
probability randomly move."""
# nothing has changed from the original implementation
rand_val = np.random.random()
q_actions = self.model.predict(data.reshape(1, self.observation_size*NUM_FRAMES), batch_size = 1)
if rand_val < epsilon:
opt_policy = np.random.randint(0, NUM_ACTIONS)
else:
opt_policy = np.argmax(np.abs(q_actions))
self.qvalue_evolution.append(q_actions[0,opt_policy])
return opt_policy, q_actions[0, opt_policy]
def train(self, s_batch, a_batch, r_batch, d_batch, s2_batch, observation_num):
"""Trains network to fit given parameters"""
# nothing has changed from the original implementation, except for changing the input dimension 'reshape'
batch_size = s_batch.shape[0]
targets = np.zeros((batch_size, NUM_ACTIONS))
for i in range(batch_size):
targets[i] = self.model.predict(s_batch[i].reshape(1, self.observation_size*NUM_FRAMES), batch_size = 1)
fut_action = self.target_model.predict(s2_batch[i].reshape(1, self.observation_size*NUM_FRAMES), batch_size = 1)
targets[i, a_batch[i]] = r_batch[i]
if d_batch[i] == False:
targets[i, a_batch[i]] += DECAY_RATE * np.max(fut_action)
loss = self.model.train_on_batch(s_batch, targets)
# Print the loss every 100 iterations.
if observation_num % 100 == 0:
print("We had a loss equal to ", loss)
def save_network(self, path):
# Saves model at specified path as h5 file
# nothing has changed
self.model.save(path)
print("Successfully saved network.")
def load_network(self, path):
# nothing has changed
self.model = load_model(path)
print("Succesfully loaded network.")
def target_train(self):
# nothing has changed from the original implementation
model_weights = self.model.get_weights()
target_model_weights = self.target_model.get_weights()
for i in range(len(model_weights)):
target_model_weights[i] = TAU * model_weights[i] + (1 - TAU) * target_model_weights[i]
self.target_model.set_weights(target_model_weights)
class DuelQ(object):
"""Constructs the desired deep q learning network"""
def __init__(self, action_size, lr=0.00001, observation_size=OBSERVATION_SIZE):
# It is not modified from Abhinav Sagar's code, except for adding the possibility to change the learning rate
# in parameter is also present the size of the action space
# (it used to be a global variable in the original code)
self.action_size = action_size
self.observation_size = observation_size
self.lr_ = lr
self.model = None
self.qvalue_evolution = []
self.construct_q_network()
def construct_q_network(self):
# Uses the network architecture found in DeepMind paper
# The inputs and outputs size have changed, as well as replacing the convolution by dense layers.
self.model = Sequential()
input_layer = Input(shape = (self.observation_size*NUM_FRAMES,))
lay1 = Dense(self.observation_size*NUM_FRAMES)(input_layer)
lay1 = Activation('relu')(lay1)
lay2 = Dense(self.observation_size)(lay1)
lay2 = Activation('relu')(lay2)
lay3 = Dense(2*NUM_ACTIONS)(lay2)
lay3 = Activation('relu')(lay3)
fc1 = Dense(NUM_ACTIONS)(lay3)
advantage = Dense(NUM_ACTIONS)(fc1)
fc2 = Dense(NUM_ACTIONS)(lay3)
value = Dense(1)(fc2)
meaner = Lambda(lambda x: K.mean(x, axis=1) )
mn_ = meaner(advantage)
tmp = subtract([advantage, mn_]) # keras doesn't like this part...
policy = add([tmp, value])
self.model = Model(inputs=[input_layer], outputs=[policy])
self.model.compile(loss='mse', optimizer=Adam(lr=self.lr_))
self.target_model = Model(inputs=[input_layer], outputs=[policy])
self.target_model.compile(loss='mse', optimizer=Adam(lr=self.lr_))
print("Successfully constructed networks.")
def predict_movement(self, data, epsilon):
"""Predict movement of game controler where is epsilon
probability randomly move."""
# only changes lie in adapting the input shape
q_actions = self.model.predict(data.reshape(1, self.observation_size*NUM_FRAMES), batch_size = 1)
opt_policy = np.argmax(q_actions)
rand_val = np.random.random()
if rand_val < epsilon:
opt_policy = np.random.randint(0, NUM_ACTIONS)
self.qvalue_evolution.append(q_actions[0,opt_policy])
return opt_policy, q_actions[0, opt_policy]
def train(self, s_batch, a_batch, r_batch, d_batch, s2_batch, observation_num):
"""Trains network to fit given parameters"""
# nothing has changed except adapting the input shapes
batch_size = s_batch.shape[0]
targets = np.zeros((batch_size, NUM_ACTIONS))
for i in range(batch_size):
targets[i] = self.model.predict(s_batch[i].reshape(1, self.observation_size*NUM_FRAMES), batch_size = 1)
fut_action = self.target_model.predict(s2_batch[i].reshape(1, self.observation_size*NUM_FRAMES), batch_size = 1)
targets[i, a_batch[i]] = r_batch[i]
if d_batch[i] == False:
targets[i, a_batch[i]] += DECAY_RATE * np.max(fut_action)
loss = self.model.train_on_batch(s_batch, targets)
# Print the loss every 100 iterations.
if observation_num % 100 == 0:
print("We had a loss equal to ", loss)
def save_network(self, path):
# Saves model at specified path as h5 file
# nothing has changed
self.model.save(path)
print("Successfully saved network.")
def load_network(self, path):
# nothing has changed
self.model.load_weights(path)
self.target_model.load_weights(path)
print("Succesfully loaded network.")
def target_train(self):
# nothing has changed
model_weights = self.model.get_weights()
self.target_model.set_weights(model_weights)
Another custom made version of the Q-Learning algorithm
class RealQ(object):
"""Constructs the desired deep q learning network"""
def __init__(self, action_size, lr=1e-5, observation_size=OBSERVATION_SIZE, mean_reg=False):
self.action_size = action_size
self.observation_size = observation_size
self.model = None
self.target_model = None
self.lr_ = lr
self.mean_reg = mean_reg
self.qvalue_evolution=[]
self.construct_q_network()
def construct_q_network(self):
self.model = Sequential()
input_states = Input(shape = (self.observation_size,))
input_action = Input(shape = (NUM_ACTIONS,))
input_layer = Concatenate()([input_states, input_action])
lay1 = Dense(self.observation_size)(input_layer)
lay1 = Activation('relu')(lay1)
lay2 = Dense(self.observation_size)(lay1)
lay2 = Activation('relu')(lay2)
lay3 = Dense(2*NUM_ACTIONS)(lay2)
lay3 = Activation('relu')(lay3)
fc1 = Dense(NUM_ACTIONS)(lay3)
advantage = Dense(1, activation = 'linear')(fc1)
if self.mean_reg==True:
advantage = Lambda(lambda x : x - K.mean(x))(advantage)
self.model = Model(inputs=[input_states, input_action], outputs=[advantage])
self.model.compile(loss='mse', optimizer=Adam(lr=self.lr_))
self.model_copy = Model(inputs=[input_states, input_action], outputs=[advantage])
self.model_copy.compile(loss='mse', optimizer=Adam(lr=self.lr_))
self.model_copy.set_weights(self.model.get_weights())
def predict_movement(self, states, epsilon):
"""Predict movement of game controler where is epsilon
probability randomly move."""
# nothing has changed from the original implementation
rand_val = np.random.random()
q_actions = self.model.predict([np.tile(states.reshape(1, self.observation_size),(NUM_ACTIONS,1)), np.eye(NUM_ACTIONS)]).reshape(1,-1)
if rand_val < epsilon:
opt_policy = np.random.randint(0, NUM_ACTIONS)
else:
opt_policy = np.argmax(np.abs(q_actions))
self.qvalue_evolution.append(q_actions[0,opt_policy])
return opt_policy, q_actions[0,opt_policy]
def train(self, s_batch, a_batch, r_batch, d_batch, s2_batch, observation_num):
"""Trains network to fit given parameters"""
# nothing has changed from the original implementation, except for changing the input dimension 'reshape'
batch_size = s_batch.shape[0]
targets = np.zeros(batch_size)
last_action=np.zeros((batch_size, NUM_ACTIONS))
for i in range(batch_size):
last_action[i,a_batch[i]] = 1
q_pre = self.model.predict([s_batch[i].reshape(1, self.observation_size), last_action[i].reshape(1,-1)], batch_size=1).reshape(1,-1)
q_fut = self.model_copy.predict([np.tile(s2_batch[i].reshape(1, self.observation_size),(NUM_ACTIONS,1)), np.eye(NUM_ACTIONS)]).reshape(1,-1)
fut_action = np.max(q_fut)
if d_batch[i] == False:
targets[i] = ALPHA * (r_batch[i] + DECAY_RATE * fut_action - q_pre)
else:
targets[i] = ALPHA * (r_batch[i] - q_pre)
loss = self.model.train_on_batch([s_batch, last_action], targets)
# Print the loss every 100 iterations.
if observation_num % 100 == 0:
print("We had a loss equal to ", loss)
def save_network(self, path):
# Saves model at specified path as h5 file
# nothing has changed
self.model.save(path)
print("Successfully saved network.")
def load_network(self, path):
# nothing has changed
self.model = load_model(path)
print("Succesfully loaded network.")
def target_train(self):
# nothing has changed from the original implementation
model_weights = self.model.get_weights()
target_model_weights = self.model_copy.get_weights()
for i in range(len(model_weights)):
target_model_weights[i] = TAU * model_weights[i] + (1 - TAU) * target_model_weights[i]
self.model_copy.set_weights(model_weights)
A version of SAC algorithm as described in https://spinningup.openai.com/en/latest/algorithms/sac.html
class SAC(object):
"""Constructs the desired deep q learning network"""
def __init__(self, action_size, lr=1e-5, observation_size=OBSERVATION_SIZE):
self.action_size = action_size
self.observation_size = observation_size
self.model = None
self.target_model = None
self.lr_ = lr
self.average_reward = 0
self.life_spent = 1
self.qvalue_evolution = []
self.Is_nan = False
self.construct_q_network()
def build_q_NN(self):
model = Sequential()
input_states = Input(shape = (self.observation_size,))
input_action = Input(shape = (NUM_ACTIONS,))
input_layer = Concatenate()([input_states, input_action])
lay1 = Dense(self.observation_size)(input_layer)
lay1 = Activation('relu')(lay1)
lay2 = Dense(self.observation_size)(lay1)
lay2 = Activation('relu')(lay2)
lay3 = Dense(2*NUM_ACTIONS)(lay2)
lay3 = Activation('relu')(lay3)
advantage = Dense(1, activation = 'linear')(lay3)
model = Model(inputs=[input_states, input_action], outputs=[advantage])
model.compile(loss='mse', optimizer=Adam(lr=self.lr_))
return model
def construct_q_network(self):
#construct double Q networks
self.model_Q = self.build_q_NN()
self.model_Q2 = self.build_q_NN()
#state value function approximation
self.model_value = Sequential()
input_states = Input(shape = (self.observation_size,))
lay1 = Dense(self.observation_size)(input_states)
lay1 = Activation('relu')(lay1)
lay3 = Dense(2*NUM_ACTIONS)(lay1)
lay3 = Activation('relu')(lay3)
advantage = Dense(NUM_ACTIONS, activation = 'relu')(lay3)
state_value = Dense(1, activation = 'linear')(advantage)
self.model_value = Model(inputs=[input_states], outputs=[state_value])
self.model_value.compile(loss='mse', optimizer=Adam(lr=self.lr_))
self.model_value_target = Sequential()
input_states = Input(shape = (self.observation_size,))
lay1 = Dense(self.observation_size)(input_states)
lay1 = Activation('relu')(lay1)
lay3 = Dense(2*NUM_ACTIONS)(lay1)
lay3 = Activation('relu')(lay3)
advantage = Dense(NUM_ACTIONS, activation = 'relu')(lay3)
state_value = Dense(1, activation = 'linear')(advantage)
self.model_value_target = Model(inputs=[input_states], outputs=[state_value])
self.model_value_target.compile(loss='mse', optimizer=Adam(lr=self.lr_))
self.model_value_target.set_weights(self.model_value.get_weights())
#policy function approximation
self.model_policy = Sequential()
input_states = Input(shape = (self.observation_size,))
lay1 = Dense(self.observation_size)(input_states)
lay1 = Activation('relu')(lay1)
lay2 = Dense(self.observation_size)(lay1)
lay2 = Activation('relu')(lay2)
lay3 = Dense(2*NUM_ACTIONS)(lay2)
lay3 = Activation('relu')(lay3)
soft_proba = Dense(NUM_ACTIONS, activation="softmax", kernel_initializer='uniform')(lay3)
self.model_policy = Model(inputs=[input_states], outputs=[soft_proba])
self.model_policy.compile(loss='categorical_crossentropy', optimizer=Adam(lr=self.lr_))
print("Successfully constructed networks.")
def predict_movement(self, states, epsilon):
"""Predict movement of game controler where is epsilon
probability randomly move."""
# nothing has changed from the original implementation
p_actions = self.model_policy.predict(states.reshape(1, self.observation_size)).ravel()
rand_val = np.random.random()
if rand_val < epsilon / 10:
opt_policy = np.random.randint(0, NUM_ACTIONS)
else:
#opt_policy = np.random.choice(np.arange(NUM_ACTIONS, dtype=int), size=1, p = p_actions)
opt_policy = np.argmax(p_actions)
return np.int(opt_policy), p_actions[opt_policy]
def train(self, s_batch, a_batch, r_batch, d_batch, s2_batch, observation_num):
"""Trains networks to fit given parameters"""
# nothing has changed from the original implementation, except for changing the input dimension 'reshape'
batch_size = s_batch.shape[0]
target = np.zeros((batch_size, 1))
new_proba = np.zeros((batch_size, NUM_ACTIONS))
last_action=np.zeros((batch_size, NUM_ACTIONS))
#training of the action state value networks
last_action=np.zeros((batch_size, NUM_ACTIONS))
for i in range(batch_size):
last_action[i,a_batch[i]] = 1
v_t = self.model_value_target.predict(s_batch[i].reshape(1, self.observation_size*NUM_FRAMES), batch_size = 1)
self.qvalue_evolution.append(v_t[0])
fut_action = self.model_value_target.predict(s2_batch[i].reshape(1, self.observation_size*NUM_FRAMES), batch_size = 1)
target[i,0] = r_batch[i] + (1 - d_batch[i]) * DECAY_RATE * fut_action
loss = self.model_Q.train_on_batch([s_batch, last_action], target)
loss_2 = self.model_Q2.train_on_batch([s_batch, last_action], target)
#training of the policy
for i in range(batch_size):
self.life_spent += 1
temp = 1 / np.log(self.life_spent) / 2
new_values = self.model_Q.predict([np.tile(s_batch[i].reshape(1, self.observation_size),(NUM_ACTIONS,1)),
np.eye(NUM_ACTIONS)]).reshape(1,-1)
new_values -= np.amax(new_values, axis=-1)
new_proba[i] = np.exp(new_values / temp) / np.sum(np.exp(new_values / temp), axis=-1)
loss_policy = self.model_policy.train_on_batch(s_batch, new_proba)
#training of the value_function
value_target = np.zeros(batch_size)
for i in range(batch_size):
target_pi = self.model_policy.predict(s_batch[i].reshape(1, self.observation_size*NUM_FRAMES), batch_size = 1)
action_v1 = self.model_Q.predict([np.tile(s_batch[i].reshape(1, self.observation_size),(NUM_ACTIONS,1)),
np.eye(NUM_ACTIONS)]).reshape(1,-1)
action_v2 = self.model_Q2.predict([np.tile(s_batch[i].reshape(1, self.observation_size),(NUM_ACTIONS,1)),
np.eye(NUM_ACTIONS)]).reshape(1,-1)
value_target[i] = np.fmin(action_v1[0,a_batch[i]], action_v2[0,a_batch[i]]) - np.sum(target_pi * np.log(target_pi + 1e-6))
loss_value = self.model_value.train_on_batch(s_batch, value_target.reshape(-1,1))
self.Is_nan = np.isnan(loss) + np.isnan(loss_2) + np.isnan(loss_policy) + np.isnan(loss_value)
# Print the loss every 100 iterations.
if observation_num % 100 == 0:
print("We had a loss equal to ", loss, loss_2, loss_policy, loss_value)
def save_network(self, path):
# Saves model at specified path as h5 file
# nothing has changed
self.model_policy.save("policy_"+path)
self.model_value_target.save("value_"+path)
print("Successfully saved network.")
def load_network(self, path):
# nothing has changed
self.model_policy = load_model("policy_"+path)
elf.model_value_target = load_model("value_"+path)
print("Succesfully loaded network.")
def target_train(self):
# nothing has changed from the original implementation
model_weights = self.model_value.get_weights()
target_model_weights = self.model_value_target.get_weights()
for i in range(len(model_weights)):
target_model_weights[i] = TAU * model_weights[i] + (1 - TAU) * target_model_weights[i]
self.model_value_target.set_weights(model_weights)
We first show how the agent code looks like without the "utilities" to train it. As we can see, defining this agent is pretty simple. The only real method that need to be adapted it "my_act
" that is a method of 2 lines of code.
# Credit Abhinav Sagar:
# https://github.com/abhinavsagar/Reinforcement-Learning-Tutorial
# Code under MIT license, available at:
# https://github.com/abhinavsagar/Reinforcement-Learning-Tutorial/blob/master/LICENSE
from grid2op.Parameters import Parameters
from grid2op.Agent import AgentWithConverter
from grid2op.Converter import IdToAct
import pdb
class DeepQAgent(AgentWithConverter):
# first change: An Agent must derived from grid2op.Agent (in this case MLAgent, because we manipulate vector instead
# of classes)
def convert_obs(self, observation):
return observation.to_vect()
def my_act(self, transformed_observation, reward, done=False):
if self.deep_q is None:
self.init_deep_q(transformed_observation)
predict_movement_int, *_ = self.deep_q.predict_movement(transformed_observation, epsilon=0.0)
# print("predict_movement_int: {}".format(predict_movement_int))
return predict_movement_int
def init_deep_q(self, transformed_observation):
if self.deep_q is None:
# the first time an observation is observed, I set up the neural network with the proper dimensions.
if self.mode == "DQN":
cls = DeepQ
elif self.mode == "DDQN":
cls = DuelQ
elif self.mode == "RealQ":
cls = RealQ
elif self.mode == "SAC":
cls = SAC
else:
raise RuntimeError("Unknown neural network named \"{}\"".format(self.mode))
self.deep_q = cls(self.action_space.size(), observation_size=transformed_observation.shape[0], lr=self.lr)
def __init__(self, action_space, mode="DDQN", lr=1e-5):
# this function has been adapted.
# to built a AgentWithConverter, we need an action_space.
# No problem, we add it in the constructor.
AgentWithConverter.__init__(self, action_space, action_space_converter=IdToAct)
# and now back to the origin implementation
self.replay_buffer = ReplayBuffer(BUFFER_SIZE)
# compare to original implementation, i don't know the observation space size.
# Because it depends on the component of the observation we want to look at. So these neural network will
# be initialized the first time an observation is observe.
self.deep_q = None
self.mode = mode
self.lr=lr
def load_network(self, path):
# not modified compare to original implementation
self.deep_q.load_network(path)
def convert_process_buffer(self):
"""Converts the list of NUM_FRAMES images in the process buffer
into one training sample"""
# here i simply concatenate the action in case of multiple action in the "buffer"
# this function existed in the original implementation, bus has been adapted.
return np.concatenate(self.process_buffer)
And now we will also define some utility class (as defined in the blog post) to make the training easier.
from grid2op.Reward import L2RPNReward
from grid2op.Reward import RedispReward
class L2RPNReward_LoadWise(L2RPNReward):
"""
Update the L2RPN reward to take into account the fact that a change in the loads sum shall not be allocated as reward for the agent.
"""
def __init__(self):
super().__init__()
def initialize(self, env):
super().initialize(env)
self.reward_min = - 10 * env.backend.n_line
self.previous_loads = self.reward_max * np.ones(env.backend.n_line)
def __call__(self, action, env, has_error, is_done, is_illegal, is_ambiguous):
if not is_done and not has_error:
line_cap = self._L2RPNReward__get_lines_capacity_usage(env)
new_loads, _, _ = env.backend.loads_info()
new_flows = np.abs(env.backend.get_line_flow())
loads_variation = (np.sum(new_loads) - np.sum(self.previous_loads)) / np.sum(self.previous_loads)
res = np.sum(line_cap + loads_variation)
else:
# no more data to consider, no powerflow has been run, reward is what it is
res = self.reward_min
return res
class L2RPNReward_LoadWise_ActionWise(L2RPNReward):
"""
Update the L2RPN reward to take into account the fact that a change in the loads sum shall not be allocated as reward for the agent.
"""
def __init__(self):
super().__init__()
def initialize(self, env):
super().initialize(env)
self.reward_min = - 10 * env.backend.n_line
self.previous_loads = self.reward_max * np.ones(env.backend.n_line)
self.last_action = env.helper_action_env({})
def __call__(self, action, env, has_error, is_done, is_illegal, is_ambiguous):
if not is_done and not has_error:
line_cap = self._L2RPNReward__get_lines_capacity_usage(env)
new_loads, _, _ = env.backend.loads_info()
new_flows = np.abs(env.backend.get_line_flow())
loads_variation = (np.sum(new_loads) - np.sum(self.previous_loads)) / np.sum(self.previous_loads)
res = np.sum(line_cap + loads_variation)
else:
# no more data to consider, no powerflow has been run, reward is what it is
res = self.reward_min
res -= (action != env.helper_action_env({})) * (action == self.last_action) * env.backend.n_line / 2
self.last_action = action
return res
class TrainAgent(object):
def __init__(self, agent, reward_fun=RedispReward, env=None):
self.agent = agent
self.reward_fun = reward_fun
self.env = env
def _build_valid_env(self):
# now we are creating a valid Environment
# it's mandatory because no environment are created when the agent is
# an Agent should not get direct access to the environment, but can interact with it only by:
# * receiving reward
# * receiving observation
# * sending action
close_env = False
if self.env is None:
self.env = grid2op.make(action_class=type(self.agent.action_space({})),
reward_class=self.reward_fun)
close_env = True
# I make sure the action space of the user and the environment are the same.
if not isinstance(self.agent.init_action_space, type(self.env.action_space)):
raise RuntimeError("Imposssible to build an agent with 2 different action space")
if not isinstance(self.env.action_space, type(self.agent.init_action_space)):
raise RuntimeError("Imposssible to build an agent with 2 different action space")
# A buffer that keeps the last `NUM_FRAMES` images
self.agent.replay_buffer.clear()
self.agent.process_buffer = []
# make sure the environment is reset
obs = self.env.reset()
self.agent.process_buffer.append(self.agent.convert_obs(obs))
do_nothing = self.env.action_space()
for _ in range(NUM_FRAMES-1):
# Initialize buffer with the first frames
s1, r1, _, _ = self.env.step(do_nothing)
self.agent.process_buffer.append(self.agent.convert_obs(s1))
return close_env
def train(self, num_frames, env=None):
# this function existed in the original implementation, but has been slightly adapted.
# first we create an environment or make sure the given environment is valid
close_env = self._build_valid_env()
# bellow that, only slight modification has been made. They are highlighted
observation_num = 0
curr_state = self.agent.convert_process_buffer()
epsilon = INITIAL_EPSILON
alive_frame = 0
total_reward = 0
while observation_num < num_frames:
if observation_num % 1000 == 999:
print(("Executing loop %d" %observation_num))
# Slowly decay the learning rate
if epsilon > FINAL_EPSILON:
epsilon -= (INITIAL_EPSILON-FINAL_EPSILON)/EPSILON_DECAY
initial_state = self.agent.convert_process_buffer()
self.agent.process_buffer = []
# it's a bit less convenient that using the SpaceInvader environment.
# first we need to initiliaze the neural network
self.agent.init_deep_q(curr_state)
# then we need to predict the next move
predict_movement_int, predict_q_value = self.agent.deep_q.predict_movement(curr_state, epsilon)
# and then we convert it to a valid action
act = self.agent.convert_act(predict_movement_int)
reward, done = 0, False
for i in range(NUM_FRAMES):
temp_observation_obj, temp_reward, temp_done, _ = self.env.step(act)
# here it has been adapted too. The observation get from the environment is
# first converted to vector
# below this line no changed have been made to the original implementation.
reward += temp_reward
self.agent.process_buffer.append(self.agent.convert_obs(temp_observation_obj))
done = done | temp_done
if done:
print("Lived with maximum time ", alive_frame)
print("Earned a total of reward equal to ", total_reward)
# reset the environment
self.env.reset()
alive_frame = 0
total_reward = 0
new_state = self.agent.convert_process_buffer()
self.agent.replay_buffer.add(initial_state, predict_movement_int, reward, done, new_state)
total_reward += reward
if self.agent.replay_buffer.size() > MIN_OBSERVATION:
s_batch, a_batch, r_batch, d_batch, s2_batch = self.agent.replay_buffer.sample(MINIBATCH_SIZE)
self.agent.deep_q.train(s_batch, a_batch, r_batch, d_batch, s2_batch, observation_num)
self.agent.deep_q.target_train()
# Save the network every 1000 iterations after 50000 iterations
if observation_num > 50000 and observation_num % 1000 == 0 and self.agent.deep_q.Is_nan == 0:
print("Saving Network")
self.agent.deep_q.save_network("saved_agent_"+self.agent.mode+".h5")
alive_frame += 1
observation_num += 1
if close_env:
print("closing env")
self.env.close()
def calculate_mean(self, num_episode = 100, env=None):
# this method has been only slightly adapted from the original implementation
# Note that it is NOT the recommended method to evaluate an Agent. Please use "Grid2Op.Runner" instead
# first we create an environment or make sure the given environment is valid
close_env = self._build_valid_env()
reward_list = []
print("Printing scores of each trial")
for i in range(num_episode):
done = False
tot_award = 0
self.env.reset()
while not done:
state = self.convert_process_buffer()
# same adapation as in "train" function.
predict_movement_int = self.agent.deep_q.predict_movement(state, 0.0)[0]
predict_movement = self.agent.convert_act(predict_movement_int)
# same adapation as in the "train" funciton
observation_obj, reward, done, _ = self.env.step(predict_movement)
observation_vect_full = observation_obj.to_vect()
tot_award += reward
self.process_buffer.append(observation)
self.process_buffer = self.process_buffer[1:]
print(tot_award)
reward_list.append(tot_award)
if close_env:
self.env.close()
return np.mean(reward_list), np.std(reward_list)
Now we can define the model (agent), and then train it.
This is done exactly the same way as in the Abhinav Sagar implementation.
NB The code bellow can take a few minutes to run. It's training a Deep Reinforcement Learning Agent afterall. It this takes too long on your machine, you can always decrease the "nb_frame", and set it to 1000 for example. In this case, the Agent will probably not be really good.
NB For a real Agent, it would take much longer to train.
nb_frame = 100
env = make_new("rte_case14_redisp", test=True)
# don't forget to set "test=False" (or remove it, as False is the default value) for "real" training
my_agent = DeepQAgent(env.action_space, mode="DDQN")
trainer = TrainAgent(agent=my_agent, env=env, reward_fun=RedispReward)
trainer.train(nb_frame)
trainer.agent.deep_q.save_network("saved_agent_"+trainer.agent.mode+".h5")
/home/donnotben/Documents/Grid2Op_dev/getting_started/grid2op/MakeEnv/MakeNew.py:138: UserWarning: You are using a development environment. This is really not recommended for training agents.
WARNING:tensorflow:From /home/donnotben/.local/lib/python3.6/site-packages/tensorflow/python/ops/resource_variable_ops.py:435: colocate_with (from tensorflow.python.framework.ops) is deprecated and will be removed in a future version. Instructions for updating: Colocations handled automatically by placer. WARNING:tensorflow:From /home/donnotben/.local/lib/python3.6/site-packages/tensorflow/python/keras/utils/losses_utils.py:170: to_float (from tensorflow.python.ops.math_ops) is deprecated and will be removed in a future version. Instructions for updating: Use tf.cast instead. Successfully constructed networks. Lived with maximum time 18 Earned a total of reward equal to 1016.9437771459925 Lived with maximum time 22 Earned a total of reward equal to 1037.8194032075908 WARNING:tensorflow:From /home/donnotben/.local/lib/python3.6/site-packages/tensorflow/python/ops/math_ops.py:3066: to_int32 (from tensorflow.python.ops.math_ops) is deprecated and will be removed in a future version. Instructions for updating: Use tf.cast instead. Lived with maximum time 31 Earned a total of reward equal to 25467.21976641859 Successfully saved network.
import matplotlib.pyplot as plt
%matplotlib inline
plt.figure(figsize=(30,20))
plt.plot(my_agent.deep_q.qvalue_evolution)
plt.axhline(y=0, linewidth=3, color='red')
plt.xlim(0, len(my_agent.deep_q.qvalue_evolution))
(0.0, 100.0)
The learning curve shown above is really poor. It's because the agent has not been trained for a long time (and because it uses a very poor input data). If you train this agent for approximately 10-12 hours, using only the relative flow (obs.rho
see the last section of this notebook for an example) you will get the following:
And now, time to test this trained agent.
To do that, we have multiple choices.
Either we recode the "DeepQAgent" class to load the stored weights (that have been saved during trainig) when it is initialized (not covered in this notebook), or we can also directly specify the "instance" of the Agent to use in the Grid2Op Runner.
To do that, it's fairly simple. First, you need to specify that you won't use the "agentClass" argument, by setting it to None
, and secondly you simply provide the agent to use in the agentInstance argument.
NB If you don't do that, the Runner will be created (the constructor will raise an exception). And if you choose to use the "agentClass" argument, your agent will be reloaded from scratch. So if it doesn't load the weights it will behave as a non trained agent, unlikely to perform well on the task.
Now that we have "successfully" trained our Agent, we will evaluating it. As opposed to the trainining, the evaluation is done classically using a standard Runner.
Note that the Runner will use a "scoring function" that might be different from the "reward function" used during training. In our case, it's not. We use the L2RPNReward
in both cases.
In the code bellow, we commented on what can be different and what must be identical for training and evaluation of model.
from grid2op.Runner import Runner
from grid2op.Chronics import GridStateFromFileWithForecasts, Multifolder
scoring_function = L2RPNReward
dict_params = env.get_params_for_runner()
dict_params["gridStateclass_kwargs"]["max_iter"] = max_iter
# make a runner from an intialized environment
runner = Runner(**dict_params)
Run the Agent and save the results. As opposed to the multiple times we exposed the "runner.run" call, we never really dive into the "path_save" argument. This path allows you to save lots of information about your Agent behaviour. Please All the informations present are shown on the documentation here.
import shutil
path_save="trained_agent_log"
# delete the previous stored results
if os.path.exists(path_save):
shutil.rmtree(path_save)
# run the episode
res = runner.run(nb_episode=1, path_save=path_save)
print("The results for the trained agent are:")
for _, chron_name, cum_reward, nb_time_step, max_ts in res:
msg_tmp = "\tFor chronics located at {}\n".format(chron_name)
msg_tmp += "\t\t - cumulative reward: {:.6f}\n".format(cum_reward)
msg_tmp += "\t\t - number of time steps completed: {:.0f} / {:.0f}".format(nb_time_step, max_ts)
print(msg_tmp)
The results for the trained agent are: For chronics located at 0 - cumulative reward: 121368.512117 - number of time steps completed: 100 / 100
Please refer to the official document for more information about the content of the directory where the data are saved. Note that the saving of the information is triggered by the "path_save" argument sent to the "runner.run" function.
Some information that will be present in this repository are:
If enabled, the :class:Runner
will save the information in a structured way. For each episode there will be a folder
with:
"episode_meta.json" that represents some meta information about:
grid2op.Backend
class usedgrid2op.Environment
class used."episode_times.json": gives some information about the total time spend in multiple part of the runner, mainly the
grid2op.Agent
(and especially its method grid2op.Agent.act
) and amount of time spent in the
grid2op.Environment
"_parameters.json": is a representation as json of a the grid2op.Parameters
used for this episode
"rewards.npy" is a numpy 1d array giving the rewards at each time step. We adopted the convention that the stored
reward at index i
is the one observed by the agent at time i
and NOT the reward sent by the
grid2op.Environment
after the action has been implemented.
"exec_times.npy" is a numpy 1d array giving the execution time of each time step of the episode
"actions.npy" gives the actions that has been taken by the grid2op.Agent
. At row i
of "actions.npy" is a
vectorized representation of the action performed by the agent at timestep i
ie. after having observed
the observation present at row i
of "observation.npy" and the reward showed in row i
of "rewards.npy".
"disc_lines.npy" gives which lines have been disconnected during the simulation of the cascading failure at each
time step. The same convention as for "rewards.npy" has been adopted. This means that the powerlines are
disconnected when the grid2op.Agent
takes the grid2op.Action
at time step i
.
"observations.npy" is a numpy 2d array reprensenting the grid2op.Observation
at the disposal of the
grid2op.Agent
when he took his action.
We can first look at the repository were the data are stored:
import os
os.listdir(path_save)
['dict_observation_space.json', 'dict_action_space.json', 'dict_env_modification_space.json', '0']
As we can see, there is only one folder there. It's named "1" because, in the original data, this came from the folder named "1" (the original data are located at "/home/donnotben/.local/lib/python3.6/site-packages/grid2op/data/test_multi_chronics/")
If there were multiple episode, each episode would have it's own folder, with a name as resemblant as possible to the origin name of the data. This is done to ease the studying of the results.
Now let's see what is inside this folder:
os.listdir(os.path.join(path_save,"0"))
['agent_exec_times.npy', 'other_rewards.json', 'env_modifications.npy', 'episode_times.json', 'actions.npy', 'observations.npy', '_parameters.json', 'rewards.npy', 'episode_meta.json', 'disc_lines_cascading_failure.npy']
We can for example load the "actions" performed by the Agent, and have a look at them.
To do that we will load the action array (represented as vector) and use the action_space to convert it back into valid action class.
all_actions = np.load(os.path.join("trained_agent_log", "0", "actions.npy"))
li_actions = []
for i in range(all_actions.shape[0]):
try:
tmp = runner.env.action_space.from_vect(all_actions[i,:])
li_actions.append(tmp)
except:
break
This allows us to have a deeper look at the action, and their effect. Note that here, we used action that can only set the line status, so looking at their effect is pretty straightforward.
Also, note that as oppose to "change", if a powerline is already connected, trying to set it as connected has absolutely no impact.
line_disc = 0
line_reco = 0
for act in li_actions:
dict_ = act.as_dict()
if "set_line_status" in dict_:
line_reco += dict_["set_line_status"]["nb_connected"]
line_disc += dict_["set_line_status"]["nb_disconnected"]
line_reco
0
As wa can see for our event, the agent always try to reconnect a powerline. As all lines are alway reconnected, this Agent does basically nothing.
We can also do the same kind of post analysis for the observation, even though here, as the observations come from files, it's probably not particularly intersting.
all_observations = np.load(os.path.join("trained_agent_log", "0", "observations.npy"))
li_observations = []
nb_real_disc = 0
for i in range(all_observations.shape[0]):
try:
tmp = runner.env.observation_space.from_vect(all_observations[i,:])
li_observations.append(tmp)
nb_real_disc += (np.sum(tmp.line_status) - tmp.line_status.shape[0])
except:
break
nb_real_disc
0
We can also look at the type of action the agent did:
actions_count = {}
for act in li_actions:
act_as_vect = tuple(act.to_vect())
if not act_as_vect in actions_count:
actions_count[act_as_vect] = 0
actions_count[act_as_vect] += 1
print("Number of different actions played: {}".format(len(actions_count)))
Number of different actions played: 1
print(runner.env.action_space.from_vect(np.array(list(actions_count.keys())[0])))
This action will: - NOT change anything to the injections - NOT perform any redispatching action - NOT force any line status - NOT switch any line status - NOT switch anything in the topology - NOT force any particular bus configuration
The agent only did one action (note that this number can really vary on the number of training step and the . This is not really good, the agent didn't learn anything.
As we saw, the agent we develop was not really interesting. To improve it, we could think about:
In this notebook, we will focus on changing the observation representation, by only feeding the agent only some informations.
To do so, the only modification we need to do is to modify the way the observation are converted. So the "convert_obs" method, and that is it. Nothing else need to be changed. Here for example we could think of only using the flow ratio (i.e., the current flows divided by the thermal limits) as part of the observation (named rho) instead of feeding the whole observation.
class DeepQAgent_Improved(DeepQAgent):
def __init__(self, action_space, mode="DDQN"):
DeepQAgent.__init__(self, action_space, mode=mode)
def convert_obs(self, observation):
return observation.rho
And we can reuse exactly the same code to train it :-)
my_agent = DeepQAgent_Improved(env.action_space, mode="DDQN")
trainer = TrainAgent(agent=my_agent, env=env)
trainer.train(nb_frame)
plt.figure(figsize=(30,20))
plt.plot(my_agent.deep_q.qvalue_evolution)
plt.axhline(y=0, linewidth=3, color='red')
_ = plt.xlim(0, len(my_agent.deep_q.qvalue_evolution))
Successfully constructed networks. Lived with maximum time 30 Earned a total of reward equal to 1037.9580153998297 Lived with maximum time 9 Earned a total of reward equal to 2160.375163460969 Lived with maximum time 32 Earned a total of reward equal to 26753.76143313748