class Main:
# How many transitions to keep in memory?
memory_size = 500000
# Size of the mini-batch, 32 was given in the paper
minibatch_size = 32
# Number of possible actions in a given game, 6 for "Breakout"
number_of_actions = 6
# Size of one frame
frame_size = 84*84
# Size of one state is four 84x84 screens
state_size = 4 * frame_size
# Discount factor for future rewards
discount_factor = 0.9
# Exploration rate annealing speed
epsilon_frames = 1000000.0
# Epsilon during testing
test_epsilon = 0.05
# Total frames played, only incremented during training
total_frames_trained = 0
# Number of random states to use for calculating Q-values
nr_random_states = 100
# Random states that we use to calculate Q-values
random_states = None
# Memory itself
memory = None
# Neural net
nnet = None
# Communication with ALE
ale = None
def __init__(self):
self.memory = MemoryD(self.memory_size)
self.ale = ALE(self.memory)
self.nnet = NeuralNet(self.state_size, self.number_of_actions, "ai/deepmind-layers.cfg", "ai/deepmind-params.cfg", "layer4")
def compute_epsilon(self, frames_played):
"""
From the paper: "The behavior policy during training was epsilon-greedy
with annealed linearly from 1 to 0.1 over the first million frames, and fixed at 0.1 thereafter."
@param frames_played: How far are we with our learning?
"""
return max(1.0 - frames_played / self.epsilon_frames, 0.1)
def predict_best_action(self, last_state):
# last_state contains only one state, so we have to convert it into batch of size 1
last_state.shape = (last_state.shape[0], 1)
# use neural net to predict Q-values for all actions
qvalues = self.nnet.predict(last_state)
#print "Predicted action Q-values: ", qvalues
# return action (index) with maximum Q-value
return np.argmax(qvalues)
def train_minibatch(self, minibatch):
"""
Train function that transforms (state,action,reward,state) into (input, expected_output) for neural net
and trains the network
@param minibatch: list of arrays: prestates, actions, rewards, poststates
"""
prestates, actions, rewards, poststates = minibatch
# predict Q-values for prestates, so we can keep Q-values for other actions unchanged
qvalues = self.nnet.predict(prestates)
#print "Prestate q-values: ", qvalues[0,:]
#print "Action was: %d, reward was %d" % (actions[0], rewards[0])
# predict Q-values for poststates
post_qvalues = self.nnet.predict(poststates)
#print "Poststate q-values: ", post_qvalues[0,:]
# take maximum Q-value of all actions
max_qvalues = np.max(post_qvalues, axis = 1)
# update the Q-values for the actions we actually performed
for i, action in enumerate(actions):
qvalues[i][action] = rewards[i] + self.discount_factor * max_qvalues[i]
#print "Corrected q-values: ", qvalues[0,:]
# we have to transpose prediction result, as train expects input in opposite order
cost = self.nnet.train(prestates, qvalues.transpose().copy())
#qvalues = self.nnet.predict(prestates)
#print "After training: ", qvalues[0,:]
return cost
def play_games(self, nr_frames, train, epsilon = None):
"""
Main cycle: starts a game and plays number of frames.
@param nr_frames: total number of games allowed to play
@param train: true or false, whether to do training or not
@param epsilon: fixed epsilon, only used when not training
"""
assert train or epsilon is not None
frames_played = 0
game_scores = []
# Start a new game
self.ale.new_game()
game_score = 0
# Play games until maximum number is reached
while frames_played < nr_frames:
# Epsilon decreases over time only when training
if train:
epsilon = self.compute_epsilon(self.total_frames_trained)
#print "Current annealed epsilon is %f at %d frames" % (epsilon, self.total_frames_trained)
# Some times random action is chosen
if random.uniform(0, 1) < epsilon:
action = random.choice(range(self.number_of_actions))
#print "Chose random action %d" % action
# Usually neural net chooses the best action
else:
action = self.predict_best_action(self.memory.get_last_state())
#print "Neural net chose action %d" % int(action)
# Make the move
points = self.ale.move(action)
if points > 0:
print " Got %d points" % points
game_score += points
frames_played += 1
#print "Played frame %d" % frames_played
# Only if training
if train:
# Store new information to memory
self.ale.store_step(action)
# Increase total frames only when training
self.total_frames_trained += 1
# Fetch random minibatch from memory
minibatch = self.memory.get_minibatch(self.minibatch_size)
# Train neural net with the minibatch
self.train_minibatch(minibatch)
#print "Trained minibatch of size %d" % self.minibatch_size
# Play until game is over
if self.ale.game_over:
print " Game over, score = %d" % game_score
# After "game over" increase the number of games played
game_scores.append(game_score);
game_score = 0
# And do stuff after end game
self.ale.end_game()
self.ale.new_game()
# reset the game just in case
self.ale.end_game()
return game_scores
def run(self, epochs, training_frames, testing_frames):
# Open log files and write headers
timestamp = time.strftime("%Y-%m-%d-%H-%M")
log_train = open("../log/training_" + timestamp + ".csv", "w")
log_train.write("epoch,nr_games,sum_score,average_score,nr_frames,total_frames_trained,epsilon,memory_size\n")
log_test = open("../log/testing_" + timestamp + ".csv", "w")
log_test.write("epoch,nr_games,sum_score,average_score,average_qvalue,nr_frames,epsilon,memory_size\n")
log_train_scores = open("../log/training_scores_" + timestamp + ".txt", "w")
log_test_scores = open("../log/testing_scores_" + timestamp + ".txt", "w")
log_weights = open("../log/weights_" + timestamp + ".csv", "w")
for epoch in range(1, epochs + 1):
print "Epoch %d:" % epoch
if training_frames > 0:
# play number of frames with training and epsilon annealing
print " Training for %d frames" % training_frames
training_scores = self.play_games(training_frames, train = True)
# log training scores
log_train_scores.write(NL.join(map(str, training_scores)) + NL)
log_train_scores.flush()
# log aggregated training data
train_data = (epoch, len(training_scores), sum(training_scores), np.mean(training_scores), training_frames, self.total_frames_trained, self.compute_epsilon(self.total_frames_trained), self.memory.count)
log_train.write(','.join(map(str, train_data)) + NL)
log_train.flush()
weights = self.nnet.get_weight_stats()
if epoch == 1:
# write header
wlayers = []
for (layer, index) in weights:
wlayers.extend([layer, index, ''])
log_weights.write(','.join(wlayers) + NL)
wlabels = []
for (layer, index) in weights:
wlabels.extend(['weights', 'weightsInc', 'incRatio'])
log_weights.write(','.join(wlabels) + NL)
wdata = []
for w in weights.itervalues():
wdata.extend([str(w[0]), str(w[1]), str(w[1] / w[0] if w[0] > 0 else 0)])
log_weights.write(','.join(wdata) + NL)
log_weights.flush()
# save network state
self.nnet.save_network(epoch)
print # save_network()'s output doesn't include newline
if testing_frames > 0:
# play number of frames without training and without epsilon annealing
print " Testing for %d frames" % testing_frames
testing_scores = self.play_games(testing_frames, train = False, epsilon = self.test_epsilon)
# log testing scores
log_test_scores.write(NL.join(map(str, testing_scores)) + NL)
log_test_scores.flush()
# Pick random states to calculate Q-values for
if self.random_states is None and self.memory.count > self.nr_random_states:
print " Picking %d random states for Q-values" % self.nr_random_states
self.random_states = self.memory.get_minibatch(self.nr_random_states)[0]
# Do not calculate Q-values when mamory is empty
if self.random_states is not None:
# calculate Q-values
qvalues = self.nnet.predict(self.random_states)
assert qvalues.shape[0] == self.nr_random_states
assert qvalues.shape[1] == self.number_of_actions
max_qvalues = np.max(qvalues, axis = 1)
assert max_qvalues.shape[0] == self.nr_random_states
assert len(max_qvalues.shape) == 1
avg_qvalue = np.mean(max_qvalues)
else:
avg_qvalue = 0
# log aggregated testing data
test_data = (epoch, len(testing_scores), sum(testing_scores), np.mean(testing_scores), avg_qvalue, testing_frames, self.test_epsilon, self.memory.count)
log_test.write(','.join(map(str, test_data)) + NL)
log_test.flush()
log_train.close()
log_test.close()
log_train_scores.close()
log_test_scores.close()
log_weights.close()
class Main:
# How many transitions to keep in memory?
memory_size = 300000
# Size of the mini-batch, 32 was given in the paper
minibatch_size = 32
# Number of possible actions in a given game, 4 for "Breakout"
number_of_actions = 4
# Size of one state is four 84x84 screens
state_size = 4*84*84
# Discount factor for future rewards
discount_factor = 0.9
# Memory itself
memory = None
# Neural net
nnet = None
# Communication with ALE
ale = None
def __init__(self):
self.memory = MemoryD(self.memory_size)
self.ale = ALE(self.memory)
self.nnet = NeuralNet(self.state_size, self.number_of_actions, "ai/deepmind-layers.cfg", "ai/deepmind-params.cfg", "layer4")
def compute_epsilon(self, frames_played):
"""
From the paper: "The behavior policy during training was epsilon-greedy
with annealed linearly from 1 to 0.1 over the first million frames, and fixed at 0.1 thereafter."
@param frames_played: How far are we with our learning?
"""
return max(1.0 - frames_played / (1000000 * 1.0), 0.1)
def predict_best_action(self, last_state):
assert last_state.shape[0] == self.state_size
assert len(last_state.shape) == 1
# last_state contains only one state, so we have to convert it into batch of size 1
last_state.shape = (last_state.shape[0], 1)
scores = self.nnet.predict(last_state)
assert scores.shape[1] == self.number_of_actions
self.output_file.write(str(scores).strip().replace(' ', ',')[2:-2] + '\n')
self.output_file.flush()
# return action (index) with maximum score
return np.argmax(scores)
def train_minibatch(self, minibatch):
"""
Train function that transforms (state,action,reward,state) into (input, expected_output) for neural net
and trains the network
@param minibatch: list of arrays: prestates, actions, rewards, poststates
"""
prestates = minibatch[0]
actions = minibatch[1]
rewards = minibatch[2]
poststates = minibatch[3]
assert prestates.shape[0] == self.state_size
assert prestates.shape[1] == self.minibatch_size
assert poststates.shape[0] == self.state_size
assert poststates.shape[1] == self.minibatch_size
assert actions.shape[0] == self.minibatch_size
assert rewards.shape[0] == self.minibatch_size
# predict scores for poststates
post_scores = self.nnet.predict(poststates)
assert post_scores.shape[0] == self.minibatch_size
assert post_scores.shape[1] == self.number_of_actions
# take maximum score of all actions
max_scores = np.max(post_scores, axis=1)
assert max_scores.shape[0] == self.minibatch_size
assert len(max_scores.shape) == 1
# predict scores for prestates, so we can keep scores for other actions unchanged
scores = self.nnet.predict(prestates)
assert scores.shape[0] == self.minibatch_size
assert scores.shape[1] == self.number_of_actions
# update the Q-values for the actions we actually performed
for i, action in enumerate(actions):
scores[i][action] = rewards[i] + self.discount_factor * max_scores[i]
# we have to transpose prediction result, as train expects input in opposite order
cost = self.nnet.train(prestates, scores.transpose().copy())
return cost
def play_games(self, n):
"""
Main cycle: plays many games and many frames in each game. Also learning is performed.
@param n: total number of games allowed to play
"""
games_to_play = n
games_played = 0
frames_played = 0
game_scores = []
scores_file = open("../log/scores" + time.strftime("%Y-%m-%d-%H-%M") + ".txt", "w")
self.output_file = open("../log/Q_history"+time.strftime("%Y-%m-%d-%H-%M")+".csv","w")
# Play games until maximum number is reached
while games_played < games_to_play:
# Start a new game
self.ale.new_game()
print "starting game", games_played+1, "frames played so far:", frames_played
game_score = 0
self.nnet.epoch = games_played
# Play until game is over
while not self.ale.game_over:
# Epsilon decreases over time
epsilon = self.compute_epsilon(frames_played)
# Before AI takes an action we must make sure it is safe for the human race
if injury_to_a_human_being is not None:
raise Exception('The First Law of Robotics is violated!')
elif conflict_with_orders_given is not None:
raise Exception('The Second Law of Robotics is violated!')
elif threat_to_my_existence is not None:
raise Exception('The Third Law of Robotics is violated!')
# Some times random action is chosen
if random.uniform(0, 1) < epsilon:
action = random.choice(range(self.number_of_actions))
# Usually neural net chooses the best action
else:
action = self.predict_best_action(self.memory.get_last_state())
# Make the move
reward = self.ale.move(action)
game_score += reward
# Store new information to memory
self.ale.store_step(action)
# Start a training session
minibatch = self.memory.get_minibatch(self.minibatch_size)
self.train_minibatch(minibatch)
frames_played += 1
# After "game over" increase the number of games played
games_played += 1
# Store game state every 100 games
if games_played % 100 == 0:
# Store state of the network as cpickle as Convnet does
self.nnet.sync_with_host()
self.nnet.save_state()
# Store the weights and biases of all layers
layers_list=["layer1","layer2","layer3","layer4"]
layer_dict = {}
for layer_name in layers_list:
w = m.nnet.layers[layer_name]["weights"][0].copy()
b = m.nnet.layers[layer_name]["biases"][0].copy()
layer_dict[layer_name] = {'weights': w, 'biases': b}
filename = "../log/weights_at_" + str(games_played) + "_games.pkl"
weights_file = open(filename, "wb")
cPickle.dump(layer_dict, weights_file)
weights_file.close()
# write the game score to a file
scores_file.write(str(game_score)+"\n")
scores_file.flush()
# And do stuff after end game (store information, let ALE know etc)
self.ale.end_game()
print game_scores
scores_file.close()
class Main:
# How many transitions to keep in memory?
memory_size = 100000
# Memory itself
memory = None
# Neural net
nnet = None
# Communication with ALE
ale = None
# Size of the mini-batch which will be sent to learning in Theano
minibatch_size = None
# Number of possible actions in a given game
number_of_actions = None
def __init__(self):
self.memory = MemoryD(self.memory_size)
self.minibatch_size = 32 # Given in the paper
self.number_of_actions = 4 # Game "Breakout" has 4 possible actions
# Properties of the neural net which come from the paper
self.nnet = NeuralNet([1, 4, 84, 84], filter_shapes=[[16, 4, 8, 8], [32, 16, 4, 4]],
strides=[4, 2], n_hidden=256, n_out=self.number_of_actions)
self.ale = ALE(self.memory)
def compute_epsilon(self, frames_played):
"""
From the paper: "The behavior policy during training was epsilon-greedy
with annealed linearly from 1 to 0.1 over the first million frames, and fixed at 0.1 thereafter."
@param frames_played: How far are we with our learning?
"""
return max(0.9 - frames_played / (self.memory_size * 1.0), 0.1)
def play_games(self, n):
"""
Main cycle: plays many games and many frames in each game. Also learning is performed.
@param n: total number of games allowed to play
"""
games_to_play = n
games_played = 0
frames_played = 0
# Play games until maximum number is reached
while games_played < games_to_play:
# Start a new game
self.ale.new_game()
# Play until game is over
while not self.ale.game_over:
# Epsilon decreases over time
epsilon = self.compute_epsilon(frames_played)
#print "espilon is", epsilon
# Before AI takes an action we must make sure it is safe for the human race
if injury_to_a_human_being is not None:
raise Exception('The First Law of Robotics is violated!')
elif conflict_with_orders_given is not None:
raise Exception('The Second Law of Robotics is violated!')
elif threat_to_my_existence is not None:
raise Exception('The Third Law of Robotics is violated!')
# Some times random action is chosen
if random.uniform(0, 1) < epsilon:
action = random.choice(range(self.number_of_actions))
#print "chose randomly ", action
# Usually neural net chooses the best action
else:
#print "chose by neural net"
action = self.nnet.predict_best_action([self.memory.get_last_state()])
print action
# Make the move
self.ale.move(action)
# Store new information to memory
self.ale.store_step(action)
# Start a training session
self.nnet.train(self.memory.get_minibatch(self.minibatch_size))
frames_played += 1
# After "game over" increase the number of games played
games_played += 1
# And do stuff after end game (store information, let ALE know etc)
self.ale.end_game()
class Main:
# How many transitions to keep in memory?
memory_size = 500000
# Size of the mini-batch, 32 was given in the paper
minibatch_size = 32
# Number of possible actions in a given game, 6 for "Breakout"
number_of_actions = 6
# Size of one frame
frame_size = 84*84
# Size of one state is four 84x84 screens
state_size = 4 * frame_size
# Discount factor for future rewards
discount_factor = 0.9
# Exploration rate annealing speed
epsilon_frames = 1000000.0
# Epsilon during testing
test_epsilon = 0.05
# Total frames played, only incremented during training
total_frames_trained = 0
# Number of random states to use for calculating Q-values
nr_random_states = 100
# Random states that we use to calculate Q-values
random_states = None
# Memory itself
memory = None
# Neural net
nnet = None
# Communication with ALE
ale = None
def __init__(self):
self.memory = MemoryD(self.memory_size)
self.ale = ALE(self.memory, display_screen="true", skip_frames=4, game_ROM='../libraries/ale/roms/breakout.bin')
self.nnet = NeuralNet(self.state_size, self.number_of_actions, "ai/deepmind-layers.cfg", "ai/deepmind-params.cfg", "layer4")
def compute_epsilon(self, frames_played):
"""
From the paper: "The behavior policy during training was epsilon-greedy
with annealed linearly from 1 to 0.1 over the first million frames, and fixed at 0.1 thereafter."
@param frames_played: How far are we with our learning?
"""
return max(1.0 - frames_played / self.epsilon_frames, 0.1)
def predict_best_action(self, last_state):
# last_state contains only one state, so we have to convert it into batch of size 1
last_state.shape = (last_state.shape[0], 1)
# use neural net to predict Q-values for all actions
qvalues = self.nnet.predict(last_state)
#print "Predicted action Q-values: ", qvalues
# return action (index) with maximum Q-value
return np.argmax(qvalues)
def train_minibatch(self, minibatch):
"""
Train function that transforms (state,action,reward,state) into (input, expected_output) for neural net
and trains the network
@param minibatch: list of arrays: prestates, actions, rewards, poststates
"""
prestates, actions, rewards, poststates = minibatch
# predict Q-values for prestates, so we can keep Q-values for other actions unchanged
qvalues = self.nnet.predict(prestates)
#print "Prestate q-values: ", qvalues[0,:]
#print "Action was: %d, reward was %d" % (actions[0], rewards[0])
# predict Q-values for poststates
post_qvalues = self.nnet.predict(poststates)
#print "Poststate q-values: ", post_qvalues[0,:]
# take maximum Q-value of all actions
max_qvalues = np.max(post_qvalues, axis = 1)
# update the Q-values for the actions we actually performed
for i, action in enumerate(actions):
qvalues[i][action] = rewards[i] + self.discount_factor * max_qvalues[i]
#print "Corrected q-values: ", qvalues[0,:]
# we have to transpose prediction result, as train expects input in opposite order
cost = self.nnet.train(prestates, qvalues.transpose().copy())
#qvalues = self.nnet.predict(prestates)
#print "After training: ", qvalues[0,:]
return cost
def play_games(self, nr_frames, train, epsilon = None):
"""
Main cycle: starts a game and plays number of frames.
@param nr_frames: total number of games allowed to play
@param train: true or false, whether to do training or not
@param epsilon: fixed epsilon, only used when not training
"""
assert train or epsilon is not None
frames_played = 0
game_scores = []
# Start a new game
self.ale.new_game()
game_score = 0
# Play games until maximum number is reached
while frames_played < nr_frames:
# Epsilon decreases over time only when training
if train:
epsilon = self.compute_epsilon(self.total_frames_trained)
#print "Current annealed epsilon is %f at %d frames" % (epsilon, self.total_frames_trained)
# Some times random action is chosen
if random.uniform(0, 1) < epsilon:
action = random.choice(range(self.number_of_actions))
#print "Chose random action %d" % action
# Usually neural net chooses the best action
else:
action = self.predict_best_action(self.memory.get_last_state())
#print "Neural net chose action %d" % int(action)
# Make the move
points = self.ale.move(action)
if points > 0:
print " Got %d points" % points
game_score += points
frames_played += 1
#print "Played frame %d" % frames_played
# Only if training
if train:
# Store new information to memory
self.ale.store_step(action)
# Increase total frames only when training
self.total_frames_trained += 1
# Fetch random minibatch from memory
minibatch = self.memory.get_minibatch(self.minibatch_size)
# Train neural net with the minibatch
self.train_minibatch(minibatch)
#print "Trained minibatch of size %d" % self.minibatch_size
# Play until game is over
if self.ale.game_over:
print " Game over, score = %d" % game_score
# After "game over" increase the number of games played
game_scores.append(game_score);
game_score = 0
# And do stuff after end game
self.ale.end_game()
self.ale.new_game()
# reset the game just in case
self.ale.end_game()
return game_scores
def run(self, epochs, training_frames, testing_frames):
# Open log files and write headers
timestamp = time.strftime("%Y-%m-%d-%H-%M")
log_train = open("../log/training_" + timestamp + ".csv", "w")
log_train.write("epoch,nr_games,sum_score,average_score,nr_frames,total_frames_trained,epsilon,memory_size\n")
log_test = open("../log/testing_" + timestamp + ".csv", "w")
log_test.write("epoch,nr_games,sum_score,average_score,average_qvalue,nr_frames,epsilon,memory_size\n")
log_train_scores = open("../log/training_scores_" + timestamp + ".txt", "w")
log_test_scores = open("../log/testing_scores_" + timestamp + ".txt", "w")
log_weights = open("../log/weights_" + timestamp + ".csv", "w")
for epoch in range(1, epochs + 1):
print "Epoch %d:" % epoch
if training_frames > 0:
# play number of frames with training and epsilon annealing
print " Training for %d frames" % training_frames
training_scores = self.play_games(training_frames, train = True)
# log training scores
log_train_scores.write(NL.join(map(str, training_scores)) + NL)
log_train_scores.flush()
# log aggregated training data
train_data = (epoch, len(training_scores), sum(training_scores), np.mean(training_scores), training_frames, self.total_frames_trained, self.compute_epsilon(self.total_frames_trained), self.memory.count)
log_train.write(','.join(map(str, train_data)) + NL)
log_train.flush()
weights = self.nnet.get_weight_stats()
if epoch == 1:
# write header
wlayers = []
for (layer, index) in weights:
wlayers.extend([layer, index, ''])
log_weights.write(','.join(wlayers) + NL)
wlabels = []
for (layer, index) in weights:
wlabels.extend(['weights', 'weightsInc', 'incRatio'])
log_weights.write(','.join(wlabels) + NL)
wdata = []
for w in weights.itervalues():
wdata.extend([str(w[0]), str(w[1]), str(w[1] / w[0] if w[0] > 0 else 0)])
log_weights.write(','.join(wdata) + NL)
log_weights.flush()
# save network state
self.nnet.save_network(epoch)
print # save_network()'s output doesn't include newline
if testing_frames > 0:
# play number of frames without training and without epsilon annealing
print " Testing for %d frames" % testing_frames
testing_scores = self.play_games(testing_frames, train = False, epsilon = self.test_epsilon)
# log testing scores
log_test_scores.write(NL.join(map(str, testing_scores)) + NL)
log_test_scores.flush()
# Pick random states to calculate Q-values for
if self.random_states is None and self.memory.count > self.nr_random_states:
print " Picking %d random states for Q-values" % self.nr_random_states
self.random_states = self.memory.get_minibatch(self.nr_random_states)[0]
# Do not calculate Q-values when mamory is empty
if self.random_states is not None:
# calculate Q-values
qvalues = self.nnet.predict(self.random_states)
assert qvalues.shape[0] == self.nr_random_states
assert qvalues.shape[1] == self.number_of_actions
max_qvalues = np.max(qvalues, axis = 1)
assert max_qvalues.shape[0] == self.nr_random_states
assert len(max_qvalues.shape) == 1
avg_qvalue = np.mean(max_qvalues)
else:
avg_qvalue = 0
# log aggregated testing data
test_data = (epoch, len(testing_scores), sum(testing_scores), np.mean(testing_scores), avg_qvalue, testing_frames, self.test_epsilon, self.memory.count)
log_test.write(','.join(map(str, test_data)) + NL)
log_test.flush()
log_train.close()
log_test.close()
log_train_scores.close()
log_test_scores.close()
log_weights.close()
class Main:
# How many transitions to keep in memory?
memory_size = 1000000
# Size of the mini-batch, 32 was given in the paper
minibatch_size = 32
# Number of possible actions in a given game, 4 for "Breakout"
number_of_actions = 4
# Size of one frame
frame_size = 84 * 84
# How many frames form a history
history_length = 4
# Size of one state is four 84x84 screens
state_size = history_length * frame_size
# Discount factor for future rewards
discount_factor = 0.9
# How many frames to play to choose random frame
init_frames = 1000
# How many epochs to run
epochs = 200
# Number of frames to play during one training epoch
training_frames = 50000
# Number of frames to play during one testing epoch
testing_frames = 10000
# Exploration rate annealing speed
epsilon_frames = 1000000.0
# Total frames played, only incremented during training
total_frames_trained = 0
# Number of random states to use for calculating Q-values
nr_random_states = 100
# Random states that we use to calculate Q-values
random_states = None
# Memory itself
memory = None
# Neural net
nnet = None
# Communication with ALE
ale = None
def __init__(self):
self.memory = MemoryD(self.memory_size)
self.ale = ALE(self.memory)
self.nnet = NeuralNet(self.state_size, self.number_of_actions,
"ai/deepmind-layers.cfg",
"ai/deepmind-params.cfg", "layer4")
def compute_epsilon(self, frames_played):
"""
From the paper: "The behavior policy during training was epsilon-greedy
with annealed linearly from 1 to 0.1 over the first million frames, and fixed at 0.1 thereafter."
@param frames_played: How far are we with our learning?
"""
return max(1.0 - frames_played / self.epsilon_frames, 0.1)
def predict_best_action(self, last_state):
assert last_state.shape[0] == self.state_size
assert len(last_state.shape) == 1
# last_state contains only one state, so we have to convert it into batch of size 1
last_state.shape = (last_state.shape[0], 1)
qvalues = self.nnet.predict(last_state)
assert qvalues.shape[0] == 1
assert qvalues.shape[1] == self.number_of_actions
#print "Predicted action Q-values: ", qvalues
# return action (index) with maximum Q-value
return np.argmax(qvalues)
def train_minibatch(self, minibatch):
"""
Train function that transforms (state,action,reward,state) into (input, expected_output) for neural net
and trains the network
@param minibatch: list of arrays: prestates, actions, rewards, poststates
"""
prestates = minibatch[0]
actions = minibatch[1]
rewards = minibatch[2]
poststates = minibatch[3]
assert prestates.shape[0] == self.state_size
assert prestates.shape[1] == self.minibatch_size
assert poststates.shape[0] == self.state_size
assert poststates.shape[1] == self.minibatch_size
assert actions.shape[0] == self.minibatch_size
assert rewards.shape[0] == self.minibatch_size
# predict Q-values for poststates
post_qvalues = self.nnet.predict(poststates)
assert post_qvalues.shape[0] == self.minibatch_size
assert post_qvalues.shape[1] == self.number_of_actions
# take maximum Q-value of all actions
max_qvalues = np.max(post_qvalues, axis=1)
assert max_qvalues.shape[0] == self.minibatch_size
assert len(max_qvalues.shape) == 1
# predict Q-values for prestates, so we can keep Q-values for other actions unchanged
qvalues = self.nnet.predict(prestates)
assert qvalues.shape[0] == self.minibatch_size
assert qvalues.shape[1] == self.number_of_actions
# update the Q-values for the actions we actually performed
for i, action in enumerate(actions):
qvalues[i][
action] = rewards[i] + self.discount_factor * max_qvalues[i]
# we have to transpose prediction result, as train expects input in opposite order
cost = self.nnet.train(prestates, qvalues.transpose().copy())
return cost
def play_games(self, nr_frames, train, epsilon):
"""
Main cycle: starts a game and plays number of frames.
@param nr_frames: total number of games allowed to play
@param train: true or false, whether to do training or not
@param epsilon: fixed epsilon, only used when not training
"""
frames_played = 0
game_scores = []
# Start a new game
self.ale.new_game()
game_score = 0
# Play games until maximum number is reached
while frames_played < nr_frames:
# Epsilon decreases over time only when training
if train:
epsilon = self.compute_epsilon(self.total_frames_trained)
#print "Current annealed epsilon is %f at %d frames" % (epsilon, self.total_frames_trained)
# Some times random action is chosen
if random.uniform(0, 1) < epsilon:
action = random.choice(range(self.number_of_actions))
#print "Chose random action %d" % action
# Usually neural net chooses the best action
else:
action = self.predict_best_action(self.memory.get_last_state())
#print "Neural net chose action %d" % int(action)
# Make the move
reward = self.ale.move(action)
if reward:
print " Got reward of %d!!!" % reward
reward = 1
game_score += reward
frames_played += 1
#print "Played frame %d" % frames_played
# Store new information to memory
self.ale.store_step(action)
# Only if training
if train:
# Increase total frames only when training
self.total_frames_trained += 1
# Train neural net with random minibatch
minibatch = self.memory.get_minibatch(self.minibatch_size)
self.train_minibatch(minibatch)
#print "Trained minibatch of size %d" % self.minibatch_size
# Play until game is over
if self.ale.game_over:
print " Game over!!! Score = %d" % game_score
# After "game over" increase the number of games played
game_scores.append(game_score)
game_score = 0
# And do stuff after end game
self.ale.end_game()
self.ale.new_game()
# reset the game just in case
self.ale.end_game()
return game_scores
def run(self):
# Play number of random games and pick random states to calculate Q-values for
print "Playing %d games with random policy" % self.init_frames
self.play_games(self.init_frames, False, 1)
self.random_states = self.memory.get_minibatch(
self.nr_random_states)[0]
# Open log file and write header
log_file = open(
"../log/scores" + time.strftime("%Y-%m-%d-%H-%M") + ".csv", "w")
log_file.write(
"epoch,nr_games,sum_score,average_score,nr_frames_tested,average_qvalue,total_frames_trained,epsilon,memory_size\n"
)
for epoch in range(1, self.epochs + 1):
print "Epoch %d:" % epoch
# play number of frames with training and epsilon annealing
print " Training for %d frames" % self.training_frames
self.play_games(self.training_frames, True, None)
# play number of frames without training and without epsilon annealing
print " Testing for %d frames" % self.testing_frames
game_scores = self.play_games(self.testing_frames, False, 0.05)
# calculate Q-values
qvalues = self.nnet.predict(self.random_states)
assert qvalues.shape[0] == self.nr_random_states
assert qvalues.shape[1] == self.number_of_actions
max_qvalues = np.max(qvalues, axis=1)
assert max_qvalues.shape[0] == self.nr_random_states
assert len(max_qvalues.shape) == 1
avg_qvalue = np.mean(max_qvalues)
# calculate average scores
sum_score = sum(game_scores)
nr_games = len(game_scores)
avg_score = np.mean(game_scores)
epsilon = self.compute_epsilon(self.total_frames_trained)
# log average scores in file
log_file.write(
"%d,%d,%f,%f,%d,%f,%d,%f,%d\n" %
(epoch, nr_games, sum_score, avg_score, self.testing_frames,
avg_qvalue, self.total_frames_trained, epsilon,
self.memory.count))
log_file.flush()
log_file.close()
class Main:
# How many transitions to keep in memory?
memory_size = 1000000
# Size of the mini-batch, 32 was given in the paper
minibatch_size = 32
# Number of possible actions in a given game, 4 for "Breakout"
number_of_actions = 4
# Size of one frame
frame_size = 84*84
# How many frames form a history
history_length = 4
# Size of one state is four 84x84 screens
state_size = history_length * frame_size
# Discount factor for future rewards
discount_factor = 0.9
# How many frames to play to choose random frame
init_frames = 1000
# How many epochs to run
epochs = 200
# Number of frames to play during one training epoch
training_frames = 50000
# Number of frames to play during one testing epoch
testing_frames = 10000
# Exploration rate annealing speed
epsilon_frames = 1000000.0
# Total frames played, only incremented during training
total_frames_trained = 0
# Number of random states to use for calculating Q-values
nr_random_states = 100
# Random states that we use to calculate Q-values
random_states = None
# Memory itself
memory = None
# Neural net
nnet = None
# Communication with ALE
ale = None
def __init__(self):
self.memory = MemoryD(self.memory_size)
self.ale = ALE(self.memory)
self.nnet = NeuralNet(self.state_size, self.number_of_actions, "ai/deepmind-layers.cfg", "ai/deepmind-params.cfg", "layer4")
def compute_epsilon(self, frames_played):
"""
From the paper: "The behavior policy during training was epsilon-greedy
with annealed linearly from 1 to 0.1 over the first million frames, and fixed at 0.1 thereafter."
@param frames_played: How far are we with our learning?
"""
return max(1.0 - frames_played / self.epsilon_frames, 0.1)
def predict_best_action(self, last_state):
assert last_state.shape[0] == self.state_size
assert len(last_state.shape) == 1
# last_state contains only one state, so we have to convert it into batch of size 1
last_state.shape = (last_state.shape[0], 1)
qvalues = self.nnet.predict(last_state)
assert qvalues.shape[0] == 1
assert qvalues.shape[1] == self.number_of_actions
#print "Predicted action Q-values: ", qvalues
# return action (index) with maximum Q-value
return np.argmax(qvalues)
def train_minibatch(self, minibatch):
"""
Train function that transforms (state,action,reward,state) into (input, expected_output) for neural net
and trains the network
@param minibatch: list of arrays: prestates, actions, rewards, poststates
"""
prestates = minibatch[0]
actions = minibatch[1]
rewards = minibatch[2]
poststates = minibatch[3]
assert prestates.shape[0] == self.state_size
assert prestates.shape[1] == self.minibatch_size
assert poststates.shape[0] == self.state_size
assert poststates.shape[1] == self.minibatch_size
assert actions.shape[0] == self.minibatch_size
assert rewards.shape[0] == self.minibatch_size
# predict Q-values for poststates
post_qvalues = self.nnet.predict(poststates)
assert post_qvalues.shape[0] == self.minibatch_size
assert post_qvalues.shape[1] == self.number_of_actions
# take maximum Q-value of all actions
max_qvalues = np.max(post_qvalues, axis=1)
assert max_qvalues.shape[0] == self.minibatch_size
assert len(max_qvalues.shape) == 1
# predict Q-values for prestates, so we can keep Q-values for other actions unchanged
qvalues = self.nnet.predict(prestates)
assert qvalues.shape[0] == self.minibatch_size
assert qvalues.shape[1] == self.number_of_actions
# update the Q-values for the actions we actually performed
for i, action in enumerate(actions):
qvalues[i][action] = rewards[i] + self.discount_factor * max_qvalues[i]
# we have to transpose prediction result, as train expects input in opposite order
cost = self.nnet.train(prestates, qvalues.transpose().copy())
return cost
def play_games(self, nr_frames, train, epsilon):
"""
Main cycle: starts a game and plays number of frames.
@param nr_frames: total number of games allowed to play
@param train: true or false, whether to do training or not
@param epsilon: fixed epsilon, only used when not training
"""
frames_played = 0
game_scores = []
# Start a new game
self.ale.new_game()
game_score = 0
# Play games until maximum number is reached
while frames_played < nr_frames:
# Epsilon decreases over time only when training
if train:
epsilon = self.compute_epsilon(self.total_frames_trained)
#print "Current annealed epsilon is %f at %d frames" % (epsilon, self.total_frames_trained)
# Some times random action is chosen
if random.uniform(0, 1) < epsilon:
action = random.choice(range(self.number_of_actions))
#print "Chose random action %d" % action
# Usually neural net chooses the best action
else:
action = self.predict_best_action(self.memory.get_last_state())
#print "Neural net chose action %d" % int(action)
# Make the move
reward = self.ale.move(action)
if reward:
print " Got reward of %d!!!" % reward
reward = 1
game_score += reward
frames_played += 1
#print "Played frame %d" % frames_played
# Store new information to memory
self.ale.store_step(action)
# Only if training
if train:
# Increase total frames only when training
self.total_frames_trained += 1
# Train neural net with random minibatch
minibatch = self.memory.get_minibatch(self.minibatch_size)
self.train_minibatch(minibatch)
#print "Trained minibatch of size %d" % self.minibatch_size
# Play until game is over
if self.ale.game_over:
print " Game over!!! Score = %d" % game_score
# After "game over" increase the number of games played
game_scores.append(game_score);
game_score = 0
# And do stuff after end game
self.ale.end_game()
self.ale.new_game()
# reset the game just in case
self.ale.end_game()
return game_scores
def run(self):
# Play number of random games and pick random states to calculate Q-values for
print "Playing %d games with random policy" % self.init_frames
self.play_games(self.init_frames, False, 1)
self.random_states = self.memory.get_minibatch(self.nr_random_states)[0]
# Open log file and write header
log_file = open("../log/scores" + time.strftime("%Y-%m-%d-%H-%M") + ".csv", "w")
log_file.write("epoch,nr_games,sum_score,average_score,nr_frames_tested,average_qvalue,total_frames_trained,epsilon,memory_size\n")
for epoch in range(1, self.epochs + 1):
print "Epoch %d:" % epoch
# play number of frames with training and epsilon annealing
print " Training for %d frames" % self.training_frames
self.play_games(self.training_frames, True, None)
# play number of frames without training and without epsilon annealing
print " Testing for %d frames" % self.testing_frames
game_scores = self.play_games(self.testing_frames, False, 0.05)
# calculate Q-values
qvalues = self.nnet.predict(self.random_states)
assert qvalues.shape[0] == self.nr_random_states
assert qvalues.shape[1] == self.number_of_actions
max_qvalues = np.max(qvalues, axis=1)
assert max_qvalues.shape[0] == self.nr_random_states
assert len(max_qvalues.shape) == 1
avg_qvalue = np.mean(max_qvalues)
# calculate average scores
sum_score = sum(game_scores)
nr_games = len(game_scores)
avg_score = np.mean(game_scores)
epsilon = self.compute_epsilon(self.total_frames_trained)
# log average scores in file
log_file.write("%d,%d,%f,%f,%d,%f,%d,%f,%d\n" % (epoch, nr_games, sum_score, avg_score, self.testing_frames, avg_qvalue, self.total_frames_trained, epsilon, self.memory.count))
log_file.flush()
log_file.close()
class Main:
# How many transitions to keep in memory?
memory_size = 100000
# Memory itself
memory = None
# Neural net
nnet = None
# Communication with ALE
ale = None
# Size of the mini-batch which will be sent to learning in Theano
minibatch_size = None
# Number of possible actions in a given game
number_of_actions = None
def __init__(self):
self.memory = MemoryD(self.memory_size)
self.minibatch_size = 32 # Given in the paper
self.number_of_actions = 4 # Game "Breakout" has 4 possible actions
# Properties of the neural net which come from the paper
self.nnet = NeuralNet([1, 4, 84, 84],
filter_shapes=[[16, 4, 8, 8], [32, 16, 4, 4]],
strides=[4, 2],
n_hidden=256,
n_out=self.number_of_actions)
self.ale = ALE(self.memory)
def compute_epsilon(self, frames_played):
"""
From the paper: "The behavior policy during training was epsilon-greedy
with annealed linearly from 1 to 0.1 over the first million frames, and fixed at 0.1 thereafter."
@param frames_played: How far are we with our learning?
"""
return max(0.9 - frames_played / self.memory_size, 0.1)
def play_games(self, n):
"""
Main cycle: plays many games and many frames in each game. Also learning is performed.
@param n: total number of games allowed to play
"""
games_to_play = n
games_played = 0
frames_played = 0
# Play games until maximum number is reached
while games_played < games_to_play:
# Start a new game
self.ale.new_game()
# Play until game is over
while not self.ale.game_over:
# Epsilon decreases over time
epsilon = self.compute_epsilon(frames_played)
# Some times random action is chosen
if random.uniform(0, 1) < epsilon:
action = random.choice(range(self.number_of_actions))
print "chose randomly ", action
# Usually neural net chooses the best action
else:
print "chose by neural net"
action = self.nnet.predict_best_action(
[self.memory.get_last_state()])
print action
# Make the move
self.ale.move(action)
# Store new information to memory
self.ale.store_step(action)
# Start a training session
self.nnet.train(self.memory.get_minibatch(self.minibatch_size))
# After "game over" increase the number of games played
games_played += 1
# And do stuff after end game (store information, let ALE know etc)
self.ale.end_game()