class Agent:
def __init__(self, model, memory=None, memory_size=500, nb_frames=None):
assert len(
model.get_output_shape_at(0)
) == 2, "Model's output shape should be (nb_samples, nb_actions)."
if memory:
self.memory = memory
else:
self.memory = ExperienceReplay(memory_size)
if not nb_frames and not model.get_input_shape_at(0)[1]:
raise Exception("Missing argument : nb_frames not provided")
elif not nb_frames:
nb_frames = model.get_input_shape_at(0)[1]
elif model.get_input_shape_at(
0
)[1] and nb_frames and model.get_input_shape_at(0)[1] != nb_frames:
raise Exception(
"Dimension mismatch : time dimension of model should be equal to nb_frames."
)
self.model = model
self.nb_frames = nb_frames
self.frames = None
@property
def memory_size(self):
return self.memory.memory_size
@memory_size.setter
def memory_size(self, value):
self.memory.memory_size = value
def reset_memory(self):
self.exp_replay.reset_memory()
def check_game_compatibility(self, game):
#if len(self.model.input_layers_node_indices) != 1:
#raise Exception('Multi node input is not supported.')
game_output_shape = (1, None) + game.get_frame().shape
if len(game_output_shape) != len(self.model.get_input_shape_at(0)):
raise Exception(
'Dimension mismatch. Input shape of the model should be compatible with the game.'
)
else:
for i in range(len(self.model.get_input_shape_at(0))):
if self.model.get_input_shape_at(0)[i] and game_output_shape[
i] and self.model.get_input_shape_at(
0)[i] != game_output_shape[i]:
raise Exception(
'Dimension mismatch. Input shape of the model should be compatible with the game.'
)
if len(
self.model.get_output_shape_at(0)
) != 2 or self.model.get_output_shape_at(0)[1] != game.nb_actions:
raise Exception(
'Output shape of model should be (nb_samples, nb_actions).')
def get_game_data(self, game):
frame = game.get_frame()
if self.frames is None:
self.frames = [frame] * self.nb_frames
else:
self.frames.append(frame)
self.frames.pop(0)
return np.expand_dims(self.frames, 0)
def clear_frames(self):
self.frames = None
def train(self,
game,
nb_epoch=1000,
batch_size=50,
gamma=0.9,
epsilon=[1., .1],
epsilon_rate=0.5,
reset_memory=False,
observe=0,
checkpoint=None,
total_sessions=0,
session_id=1):
self.check_game_compatibility(game)
ts = int(time.time())
#fn = "gold-{}.csv".format(ts)
#fn = "9nyc-250-1000-epr8-heat-adam.csv"
#fn = "400-rl-nopool.csv"
fn = "3-normal.csv"
fn2 = "heat.csv"
#advice_type = "OA"
advice_type = "OA"
meta_advice_type = "HFHA"
#meta_feedback_frequency = 0.1
#meta_feedback_frequency = 0.5 #HF!!!
meta_feedback_frequency = 0.1 #LF!!!
heatmap = [[0] * 20 for i in range(20)]
if session_id == 1:
advice_type = "OA"
if session_id == 2:
advice_type = "NA"
if session_id == 3:
advice_type = "RL"
# print(heatmap)
# with open("dummyheat.csv",'a') as f2:
# csvWriter = csv.writer(f2,delimiter=',')
# csvWriter.writerows(heatmap)
# if ( session_id >= 3 and session_id < 5 ):
# print("Switching to HFLA")
# meta_advice_type = "HFLA"
# #meta_feedback_frequency = 0.1
# elif ( session_id >= 5 and session_id < 7 ):
# print("Switching to LFHA")
# meta_feedback_frequency = 0.1
# meta_advice_type = "LFHA"
# elif ( session_id >= 7 and session_id < 9 ):
# print("Switching to LFLA")
# meta_advice_type = "LFLA"
# elif ( session_id >= 9 and session_id < 11 ):
# advice_type = "OA"
# print("Switching to NA HFLA")
# meta_advice_type = "HFLA"
# meta_feedback_frequency = 0.5
# elif ( session_id >= 11 and session_id < 13 ):
# print("Switching to NA HFLA")
# meta_advice_type = "HFLA"
# #meta_feedback_frequency = 0.1
# elif ( session_id >= 13 and session_id < 15 ):
# print("Switching to NA LFHA")
# meta_feedback_frequency = 0.1
# meta_advice_type = "LFHA"
# elif ( session_id >= 15 and session_id < 17 ):
# print("Switching to NA LFLA")
# meta_advice_type = "LFLA"
# if ( session_id >= 2 and session_id < 3 ):
# meta_feedback_frequency = 0.1
# print("Switching to LFHA")
# advice_type = "OA"
# meta_advice_type = "LFHA"
# meta_feedback_frequency = 0.1
# elif ( session_id >= 3 and session_id < 4 ):
# advice_type = "NA"
# print("Switching to NA LFHA")
# meta_feedback_frequency = 0.1
# meta_advice_type = "LFHA"
# elif ( session_id >= 4 and session_id < 5 ):
# print("Switching to NA LFLA")
# meta_feedback_frequency = 0.1
# advice_type = "NA"
# meta_advice_type = "LFLA"
# elif ( session_id >= 5 and session_id < 6 ):
# advice_type = "OA"
# print("Switching to OA HFHA")
# meta_advice_type = "HFHA"
# meta_feedback_frequency = 0.5
# elif ( session_id >= 6 and session_id < 7 ):
# advice_type = "NA"
# meta_feedback_frequency = 0.5
# print("Switching to NA HFHA")
# meta_advice_type = "HFHA"
# meta_feedback_frequency = 0.5
# elif ( session_id >= 7 and session_id < 8 ):
# advice_type = "NA"
# print("Switching to NA HFLA")
# meta_feedback_frequency = 0.5
# meta_advice_type = "HFLA"
# elif ( session_id >= 8 and session_id < 9 ):
# advice_type = "OA"
# meta_feedback_frequency = 0.5
# print("Switching to OA HFLA")
# meta_advice_type = "HFLA"
# if ( session_id >= 4 and session_id < 7 ):
# #print("Switching to LFLA")
# advice_type = "RL"
# #meta_advice_type = "LFLA"
# elif ( session_id >= 7 and session_id < 10 ):
# # with open("1RLheat.csv",'a') as f2:
# # csvWriter = csv.writer(f2,delimiter=',')
# # csvWriter.writerows(heatmap)
# # heatmap = [ [0]*20 for i in range(20)]
# advice_type = "NA"
# #print("Switching to LFHA")
# #meta_feedback_frequency = 0.1
# #meta_advice_type = "LFHA"
# elif ( session_id >= 10 ):
# # with open("1NAheat.csv",'a') as f2:
# # csvWriter = csv.writer(f2,delimiter=',')
# # csvWriter.writerows(heatmap)
# # heatmap = [ [0]*20 for i in range(20)]
# #print("Switching to LFLA")
# #meta_advice_type = "LFLA"
# advice_type = "NA"
# with open(fn,'w') as f:
# f.write('session_id,advice_type,time,epoch,frames,score,win_perc,loss'+'\n')
# f.flush()
# f.close() with open(fn,'a') as f:
with open(fn, 'a') as f:
total_frames = 0
#f.write('session_id,advice_type,time,epoch,frames,score,win_perc,loss'+'\n')
#f.flush()
if type(epsilon) in {tuple, list}:
delta = ((epsilon[0] - epsilon[1]) / (nb_epoch * epsilon_rate))
final_epsilon = epsilon[1]
epsilon = epsilon[0]
else:
final_epsilon = epsilon
model = self.model
nb_actions = model.get_output_shape_at(0)[-1]
win_count = 0
rolling_win_window = []
max_obs_loss = -99999999999999999
m_loss = -99999999
for epoch in range(nb_epoch):
lastAdviceStep = 0
adviceGiven = 0
adviceAttempts = 0
modelActions = 0
print(heatmap)
loss = 0.
game.reset()
self.clear_frames()
if reset_memory:
self.reset_memory()
game_over = False
S = self.get_game_data(game)
savedModel = False
while not game_over:
a = 0
if advice_type == "RL":
if np.random.random() < epsilon or epoch < observe:
a = int(np.random.randint(game.nb_actions))
#print("Random Action")
else:
q = model.predict(
S
) #use the prediction confidence to determine whether to ask the player for help
qs = model.predict_classes(S)
#a = int(np.argmax(qs[0]))
#highest_conf = np.amax(q)
#print("Game Grid: {}".format(game.get_grid()))
#print("Highest MSE Confidence = {}".format(highest_conf))
#a = int(np.argmax(q[0]))
a = int(np.argmax(qs[0]))
if advice_type == "OA":
if np.random.random() < epsilon or epoch < observe:
a = int(np.random.randint(game.nb_actions))
#print("Random Action")
else:
q = model.predict(
S
) #use the prediction confidence to determine whether to ask the player for help
qs = model.predict_classes(S)
#print(qs)
#print(q)
highest_loss = abs(np.amax(q)) #added ABS
lowest_loss = abs(np.amin(q))
#print(highest_loss)
#print("HighestLoss:{}".format(highest_loss))
if highest_loss > max_obs_loss and highest_loss != 0:
max_obs_loss = highest_loss
#print("MaxLoss:{}".format(highest_loss))
#inn = highest_loss / max_obs_loss
relative_cost = np.power(
lowest_loss / max_obs_loss, 0.5)
#print("RelCostA:{}".format(relative_cost))
if relative_cost < 1e-20:
relative_cost = 1e-20
relative_cost = -1 / (np.log(relative_cost) - 1)
#print("RelCostB:{}".format(relative_cost))
confidence_score_max = 1
confidence_score_min = 0.01
feedback_chance = confidence_score_min + (
confidence_score_max -
confidence_score_min) * relative_cost
if feedback_chance < 0.01:
feedback_chance = 0.01
#if feedback_chance < 0.1:
giveAdvice = False
if (random.random() < meta_feedback_frequency):
giveAdvice = True
adviceAttempts = adviceAttempts + 1
if (relative_cost <= 0.25 and game.stepsTaken >=
(lastAdviceStep + 10)) or giveAdvice == False:
#print("HC: {}".format(max_obs_loss))
modelActions = modelActions + 1
#print("Highest Loss: {} RC: {} POS: Q0:{}".format(highest_loss, relative_cost, q[0]))
a = int(np.argmax(qs[0]))
else:
if random.random() < .5 and (
meta_advice_type == "HFLA"
or meta_advice_type == "LFLA"):
lastAdviceStep = game.stepsTaken
a = int(np.random.randint(game.nb_actions))
adviceGiven = adviceGiven + 1
#print("Taking BAD Player Action")
else:
lastAdviceStep = game.stepsTaken
adviceGiven = adviceGiven + 1
x = game.location[0]
z = game.location[1]
yaw = game.location[2]
a = -1
#print(yaw)
if z <= 6:
if x < 12:
#print("Segment1")
if yaw == 270:
a = 0
if yaw == 180:
a = 1
if yaw == 90:
a = 3
if yaw == 0:
a = 2
elif x > 15:
#print("Segment2")
if yaw == 90:
a = 0
if yaw == 180:
a = 2
if yaw == 0:
a = 1
if yaw == 270:
a = 3
else:
#print("Segment3")
if yaw == 0:
a = 0
if yaw == 270:
a = 1
if yaw == 90:
a = 2
if yaw == 180:
a = 3
elif (x >= 7) and ((z == 7) or (z == 8) or
(z == 9) or (z == 10) or
(z == 11) or (z == 12)):
#print("Segment4")
if yaw == 90:
a = 0
if yaw == 180:
a = 2
if yaw == 0:
a = 1
if yaw == 270:
a = 3
elif ((x < 7) and (x > 3)) and (
(z == 7) or (z == 8) or (z == 9) or
(z == 10) or (z == 11) or (z == 12)):
if yaw == 0:
a = 0
if yaw == 270:
a = 1
if yaw == 90:
a = 2
if yaw == 180:
a = 3
elif ((x < 3)) and ((z == 7) or (z == 8) or
(z == 9) or
(z == 10) or
(z == 11) or
(z == 12)):
if yaw == 0:
a = 2
if yaw == 270:
a = 0
if yaw == 180:
a = 1
if yaw == 90:
a = 3
elif (z == 14) or (z == 15):
if yaw == 0:
a = 0
if yaw == 270:
a = 1
if yaw == 90:
a = 2
if yaw == 180:
a = 3
elif (z == 17) or (z == 16):
#print("Segment6")
if yaw == 270:
a = 0
if yaw == 180:
a = 1
if yaw == 0:
a = 2
if yaw == 90:
a = 3
elif (z > 17):
#print("Segment6")
if yaw == 270:
a = 2
if yaw == 180:
a = 0
if yaw == 0:
a = 3
if yaw == 90:
a = 1
else:
a = int(
np.random.randint(game.nb_actions))
if a == -1:
a = int(
np.random.randint(game.nb_actions))
# if z < 6 and x < 13:
# print("Segment1")
# if yaw == 270:
# a = 0
# else:
# a = 1
# elif z < 8 and x >= 13:
# print("Segment2")
# if yaw == 0:
# a = 0
# else:
# a = 1
# elif z >= 8 and x == 13:
# print("Segment3")
# if yaw == 90:
# a = 0
# else:
# a = 1
# elif z >= 8 and z<= 17 and x < 6:
# print("Segment4")
# if yaw == 0:
# a = 0
# else:
# a = 1
# elif z > 18 and x < 18:
# print("Segment5")
# if yaw == 270:
# a = 0
# else:
# a = 1
# else:
# a = int(np.argmax(q[0]))
#print("Game Grid: {}".format(game.get_grid()))
#print("Highest MSE Confidence = {}".format(highest_conf))
if advice_type == "NA":
if np.random.random() < epsilon or epoch < observe:
a = int(np.random.randint(game.nb_actions))
game.play(a)
heatmap[game.location[0]][
game.location[1]] = heatmap[game.location[0]][
game.location[1]] + 1
#f2.write('{},{},{},{}\n'.format(advice_type,game.location[0],game.location[1],1 ))
#f2.flush()
r = game.get_score()
S_prime = self.get_game_data(game)
game_over = game.is_over()
transition = [S, a, r, S_prime, game_over]
self.memory.remember(*transition)
S = S_prime
#print("Random Action")
else:
q = model.predict(
S
) #use the prediction confidence to determine whether to ask the player for help
qs = model.predict_classes(S)
highest_loss = abs(np.amax(q)) #added ABS
lowest_loss = abs(np.amin(q))
#print("HighestLoss:{}".format(highest_loss))
if highest_loss > max_obs_loss and highest_loss != 0:
max_obs_loss = highest_loss
#print("MaxLoss:{}".format(highest_loss))
#inn = highest_loss / max_obs_loss
relative_cost = np.power(
lowest_loss / max_obs_loss, 0.5)
#print("RelCostA:{}".format(relative_cost))
if relative_cost < 1e-20:
relative_cost = 1e-20
relative_cost = -1 / (np.log(relative_cost) - 1)
#print("RelCostB:{}".format(relative_cost))
confidence_score_max = 1
confidence_score_min = 0.01
feedback_chance = confidence_score_min + (
confidence_score_max -
confidence_score_min) * relative_cost
#feedback_chance = random.random()
#print("Feedback Chance: {}".format(feedback_chance))
if feedback_chance < 0.01:
feedback_chance = 0.01
#if feedback_chance > meta_feedback_frequency:
#if feedback_chance < 0.1:
#print(relative_cost)
giveAdvice = False
if (random.random() < meta_feedback_frequency):
giveAdvice = True
adviceAttempts = adviceAttempts + 1
if (relative_cost <= 0.25 and game.stepsTaken >=
(lastAdviceStep + 10)) or giveAdvice == False:
#print("Taking Model Action")
#print("HC: {}".format(max_obs_loss))
#print("Confidence: {} RC: {}".format(feedback_chance, relative_cost))
modelActions = modelActions + 1
#a = int(np.argmin(q[0]))
a = int(np.argmax(qs[0]))
game.play(a)
heatmap[game.location[0]][
game.location[1]] = heatmap[
game.location[0]][game.location[1]] + 1
#f2.write('{},{},{},{}\n'.format(advice_type,game.location[0],game.location[1],1 ))
#f2.flush()
r = game.get_score()
S_prime = self.get_game_data(game)
game_over = game.is_over()
transition = [S, a, r, S_prime, game_over]
self.memory.remember(*transition)
S = S_prime
else:
#print("Taking Player Action")
if random.random() < .5 and (
meta_advice_type == "HFLA"
or meta_advice_type == "LFLA"):
a = int(np.random.randint(game.nb_actions))
adviceGiven = adviceGiven + 1
game.play(a)
heatmap[game.location[0]][game.location[
1]] = heatmap[game.location[0]][
game.location[1]] + 1
lastAdviceStep = game.stepsTaken
#f2.write('{},{},{},{}\n'.format(advice_type,game.location[0],game.location[1],1 ))
#f2.flush()
r = game.get_score()
S_prime = self.get_game_data(game)
game_over = game.is_over()
transition = [S, a, r, S_prime, game_over]
self.memory.remember(*transition)
S = S_prime
if game_over == False:
#game.play(checkForBestMove(game.location[0],game.location[1],game.location[2]))
a = int(
np.random.randint(game.nb_actions))
game.play(a)
heatmap[game.location[0]][
game.location[1]] = heatmap[
game.location[0]][
game.location[1]] + 1
#f2.write('{},{},{},{}\n'.format(advice_type,game.location[0],game.location[1],1 ))
#f2.flush()
r = game.get_score()
S_prime = self.get_game_data(game)
game_over = game.is_over()
transition = [
S, a, r, S_prime, game_over
]
self.memory.remember(*transition)
S = S_prime
# if game_over == False:
# game.play(checkForBestMove(game.location[0],game.location[1],game.location[2]))
# heatmap[game.location[0]][game.location[1]] = heatmap[game.location[0]][game.location[1]] + 1
# #f2.write('{},{},{},{}\n'.format(advice_type,game.location[0],game.location[1],1 ))
# #f2.flush()
# r = game.get_score()
# S_prime = self.get_game_data(game)
# game_over = game.is_over()
# transition = [S, a, r, S_prime, game_over]
# self.memory.remember(*transition)
# S = S_prime
#print("Taking BAD Player Action")
else:
adviceGiven = adviceGiven + 1
lastAdviceStep = game.stepsTaken
x = game.location[0]
z = game.location[1]
yaw = game.location[2]
#print(x)
#print(z)
a = -1
#print(yaw)
if z <= 6:
if x < 12:
#print("Segment1")
if yaw == 270:
a = 0
if yaw == 180:
a = 1
if yaw == 90:
a = 3
if yaw == 0:
a = 2
elif x > 15:
#print("Segment2")
if yaw == 90:
a = 0
if yaw == 180:
a = 2
if yaw == 0:
a = 1
if yaw == 270:
a = 3
else:
#print("Segment3")
if yaw == 0:
a = 0
if yaw == 270:
a = 1
if yaw == 90:
a = 2
if yaw == 180:
a = 3
elif (x >= 7) and ((z == 7) or (z == 8) or
(z == 9) or (z == 10) or
(z == 11) or (z == 12)):
#print("Segment4")
if yaw == 90:
a = 0
if yaw == 180:
a = 2
if yaw == 0:
a = 1
if yaw == 270:
a = 3
elif ((x < 7) and (x > 3)) and (
(z == 7) or (z == 8) or (z == 9) or
(z == 10) or (z == 11) or (z == 12)):
if yaw == 0:
a = 0
if yaw == 270:
a = 1
if yaw == 90:
a = 2
if yaw == 180:
a = 3
elif ((x < 3)) and ((z == 7) or (z == 8) or
(z == 9) or
(z == 10) or
(z == 11) or
(z == 12)):
if yaw == 0:
a = 2
if yaw == 270:
a = 0
if yaw == 180:
a = 1
if yaw == 90:
a = 3
elif (z == 14) or (z == 15):
if yaw == 0:
a = 0
if yaw == 270:
a = 1
if yaw == 90:
a = 2
if yaw == 180:
a = 3
elif (z == 17) or (z == 16):
#print("Segment6")
if yaw == 270:
a = 0
if yaw == 180:
a = 1
if yaw == 0:
a = 2
if yaw == 90:
a = 3
elif (z > 17):
#print("Segment6")
if yaw == 270:
a = 2
if yaw == 180:
a = 0
if yaw == 0:
a = 3
if yaw == 90:
a = 1
else:
a = int(
np.random.randint(game.nb_actions))
if a == -1:
a = int(
np.random.randint(game.nb_actions))
# #print(yaw)
# if z < 6 and x < 13:
# #print("Segment1")
# if yaw == 270:
# a = 0
# else:
# a = 1
# elif z < 8 and x >= 13:
# #print("Segment2")
# if yaw == 0:
# a = 0
# else:
# a = 1
# elif z >= 8 and x == 13:
# #print("Segment3")
# if yaw == 90:
# a = 0
# else:
# a = 1
# elif z >= 8 and z<= 17 and x < 6:
# #print("Segment4")
# if yaw == 0:
# a = 0
# else:
# a = 1
# elif z > 18 and x < 18:
# #print("Segment5")
# if yaw == 270:
# a = 0
# else:
# a = 1
# else:
# a = int(np.argmax(q[0]))
#Play an extra 2 times (for NA friction)
game.play(a)
heatmap[game.location[0]][
game.location[1]] = heatmap[
game.location[0]][game.location[1]] + 1
#f2.write('{},{},{},{}\n'.format(advice_type,game.location[0],game.location[1],1 ))
#f2.flush()
r = game.get_score()
S_prime = self.get_game_data(game)
game_over = game.is_over()
transition = [S, a, r, S_prime, game_over]
self.memory.remember(*transition)
S = S_prime
if game_over == False:
game.play(
checkForBestMove(
game.location[0], game.location[1],
game.location[2]))
heatmap[game.location[0]][game.location[
1]] = heatmap[game.location[0]][
game.location[1]] + 1
#f2.write('{},{},{},{}\n'.format(advice_type,game.location[0],game.location[1],1 ))
#f2.flush()
r = game.get_score()
S_prime = self.get_game_data(game)
game_over = game.is_over()
transition = [S, a, r, S_prime, game_over]
self.memory.remember(*transition)
S = S_prime
# if game_over == False:
# game.play(checkForBestMove(game.location[0],game.location[1],game.location[2]))
# heatmap[game.location[0]][game.location[1]] = heatmap[game.location[0]][game.location[1]] + 1
# #f2.write('{},{},{},{}\n'.format(advice_type,game.location[0],game.location[1],1 ))
# #f2.flush()
# r = game.get_score()
# S_prime = self.get_game_data(game)
# game_over = game.is_over()
# transition = [S, a, r, S_prime, game_over]
# self.memory.remember(*transition)
# S = S_prime
if game_over == False:
if advice_type != "NA":
game.play(a)
heatmap[game.location[0]][
game.location[1]] = heatmap[game.location[0]][
game.location[1]] + 1
#f2.write('{},{},{},{}\n'.format(advice_type,game.location[0],game.location[1],1 ))
#f2.flush()
r = game.get_score()
S_prime = self.get_game_data(game)
game_over = game.is_over()
transition = [S, a, r, S_prime, game_over]
self.memory.remember(*transition)
S = S_prime
if epoch >= observe:
batch = self.memory.get_batch(model=model,
batch_size=batch_size,
gamma=gamma)
if batch:
inputs, targets = batch
mtob = model.train_on_batch(inputs, targets)
if mtob > m_loss:
m_loss = mtob
loss += float(mtob)
#print( "LOSS: {} CULM_LOSS: {}".format(mtob,loss))
if checkpoint and (savedModel == False) and (
(epoch + 1 - observe) % checkpoint == 0
or epoch + 1 == nb_epoch):
#model.save_weights('weights.dat')
print("Checkpoint... saving model..")
if advice_type == "OA":
model.save('oa_model.h5')
if advice_type == "NA":
model.save('na_model.h5')
if advice_type == "RL":
model.save('rl_model.h5')
# model_json = model.to_json()
# with open("model.json", "w") as json_file:
# json_file.write(model_json)
# #serialize weights to HDF5
# model.save_weights("model.h5")
savedModel = True
if game.is_won():
win_count += 1
rolling_win_window.insert(0, 1)
else:
rolling_win_window.insert(0, 0)
if epsilon > final_epsilon and epoch >= observe:
epsilon -= delta
percent_win = 0
cdt = datetime.datetime.now()
if sum(rolling_win_window) != 0:
percent_win = sum(rolling_win_window) / 4
total_frames = total_frames + game.stepsTaken
f.write(
'{},{},{},{},{},{},{},{},{},{},{},{},{},{}\n'.format(
session_id, advice_type, meta_advice_type,
str(cdt), (epoch + 1), total_frames, game.score,
percent_win, epsilon, loss, game.stepsTaken,
adviceGiven, adviceAttempts, modelActions))
f.flush()
print(
"Session: {} | Time: {} | Epoch {:03d}/{:03d} | Steps {:.4f} | Epsilon {:.2f} | Score {} | Loss {}"
.format(session_id, str(cdt), epoch + 1, nb_epoch,
game.stepsTaken, epsilon, game.score, loss))
if len(rolling_win_window) > 4:
rolling_win_window.pop()
time.sleep(1.0)
if advice_type == "OA":
with open("{}OAheatxtues.csv".format(session_id), 'w+') as f2:
csvWriter = csv.writer(f2, delimiter=',')
csvWriter.writerows(heatmap)
#heatmap = [ [0]*20 for i in range(20)]
if advice_type == "RL":
with open("{}RLheatxtues.csv".format(session_id), 'w+') as f2:
csvWriter = csv.writer(f2, delimiter=',')
csvWriter.writerows(heatmap)
#heatmap = [ [0]*20 for i in range(20)]
if advice_type == "NA":
with open("{}NAheatxtues.csv".format(session_id), 'w+') as f2:
csvWriter = csv.writer(f2, delimiter=',')
csvWriter.writerows(heatmap)
#heatmap = [ [0]*20 for i in range(20)]
def play(self, game, nb_epoch=10, epsilon=0., visualize=False):
self.check_game_compatibility(game)
model = self.model
win_count = 0
frames = []
for epoch in range(nb_epoch):
print("Playing")
game.reset()
self.clear_frames()
S = self.get_game_data(game)
if visualize:
frames.append(game.draw())
game_over = False
while not game_over:
if np.random.rand() < epsilon:
print("random")
action = int(np.random.randint(0, game.nb_actions))
else:
q = model.predict(S)[0]
possible_actions = game.get_possible_actions()
q = [q[i] for i in possible_actions]
action = possible_actions[np.argmax(q)]
print(action)
game.play(action)
S = self.get_game_data(game)
if visualize:
frames.append(game.draw())
game_over = game.is_over()
if game.is_won():
win_count += 1
print("Accuracy {} %".format(100. * win_count / nb_epoch))
#Visualizing/printing images is currently super slow
if visualize:
if 'images' not in os.listdir('.'):
os.mkdir('images')
for i in range(len(frames)):
plt.imshow(frames[i], interpolation='none')
plt.savefig("images/" + game.name + str(i) + ".png")
class Agent:
def __init__(self, model, memory=None, memory_size=100, nb_frames=None):
assert len(
model.output_shape
) == 2, "Model's output shape should be (nb_samples, nb_actions)."
if memory:
self.memory = memory
else:
self.memory = ExperienceReplay(memory_size)
if not nb_frames and not model.input_shape[1]:
raise Exception("Missing argument : nb_frames not provided")
elif not nb_frames:
nb_frames = model.input_shape[1]
elif model.input_shape[
1] and nb_frames and model.input_shape[1] != nb_frames:
raise Exception(
"Dimension mismatch : time dimension of model should be equal to nb_frames."
)
self.model = model
self.nb_frames = nb_frames # model input shape, 24
self.frames = None
@property
def memory_size(self):
return self.memory.memory_size
@memory_size.setter
def memory_size(self, value):
self.memory.memory_size = value
def reset_memory(self):
self.exp_replay.reset_memory()
def check_game_compatibility(self, game):
game_output_shape = (1, None) + game.get_frame().shape
#game_output_shape = (None, game.get_frame().shape)
if len(game_output_shape) != len(self.model.input_shape):
raise Exception(
'Dimension mismatch. Input shape of the model should be compatible with the game.'
)
else:
for i in range(len(self.model.input_shape)):
if self.model.input_shape[i] and game_output_shape[
i] and self.model.input_shape[i] != game_output_shape[
i]:
raise Exception(
'Dimension mismatch. Input shape of the model should be compatible with the game.'
)
if len(self.model.output_shape
) != 2 or self.model.output_shape[1] != game.nb_actions:
raise Exception(
'Output shape of model should be (nb_samples, nb_actions).')
def get_game_data(self, game): # returns scaled
frame = game.get_frame() # candidate to return scaled
if self.frames is None:
self.frames = [frame] * self.nb_frames
else:
self.frames.append(frame)
self.frames.pop(0)
return np.expand_dims(self.frames, 0)
def clear_frames(self):
self.frames = None
def train(self,
game,
nb_epoch=1000,
batch_size=50,
gamma=0.9,
epsilon=[1., .1],
epsilon_rate=0.5,
reset_memory=False,
observe=0,
checkpoint=None):
self.check_game_compatibility(game)
if type(epsilon) in {tuple, list}:
delta = ((epsilon[0] - epsilon[1]) / (nb_epoch * epsilon_rate))
final_epsilon = epsilon[1]
epsilon = epsilon[0]
else:
final_epsilon = epsilon
save = Save()
model = self.model
nb_actions = model.output_shape[-1]
win_count = 0
for epoch in range(nb_epoch):
loss = 0.
q = np.zeros(3)
game.reset()
self.clear_frames()
if reset_memory:
self.reset_memory()
game_over = False
S = self.get_game_data(game) # S must be scaled
i = 0
while not game_over:
i = i + 1
if np.random.random() < epsilon or epoch < observe:
a = int(np.random.randint(game.nb_actions))
print('>'),
else:
# S must be scaled
q = model.predict(S) # !
a = int(np.argmax(q[0]))
game.play(a)
r = game.get_score(a)
S_prime = self.get_game_data(game) # S_prime must be scaled
game_over = game.is_over()
# S, a, S_prime, must be scaled
# reward, game over is not scaled in catch/snake
transition = [S, a, r, S_prime, game_over] # !
self.memory.remember(*transition)
S = S_prime
if epoch >= observe:
batch = self.memory.get_batch(model=model,
batch_size=batch_size,
gamma=gamma)
if batch:
inputs, targets = batch # scaled
loss += float(model.train_on_batch(inputs, targets))
#if checkpoint and ((epoch + 1 - observe) % checkpoint == 0 or epoch + 1 == nb_epoch):
#model.save_weights('4kweights.dat')
save.log(game, epoch)
if game.is_won():
win_count += 1
if epsilon > final_epsilon and epoch >= observe:
epsilon -= delta
print(' ')
print(
"Epoch {:03d}/{:03d} | Loss {:.4f} | Epsilon {:.2f} | Win count {} | loss Avg {:.4f}"
.format(epoch + 1, nb_epoch, loss, epsilon, win_count,
loss / i))
if ((epoch % 10) == 0):
save.save_model(model, Config.f_model)
save.log_epoch(loss, win_count, loss / i)
def play(self, game, nb_epoch=10, epsilon=0., visualize=True):
self.check_game_compatibility(game)
model = self.model
win_count = 0
frames = []
save = Save()
for epoch in range(nb_epoch):
game.reset()
self.clear_frames()
S = self.get_game_data(game) # S must be scaled
if visualize:
frames.append(game.draw())
game_over = False
while not game_over:
if np.random.rand() < epsilon:
print("random")
action = int(np.random.randint(0, game.nb_actions))
else:
# S must be scaled
q = model.predict(S)[0] # !
possible_actions = game.get_possible_actions()
q = [q[i] for i in possible_actions]
action = possible_actions[np.argmax(q)]
game.play(action)
S = self.get_game_data(game)
'''
if visualize:
frames.append(game.draw())
game_over = game.is_over()
'''
save.log(game, nb_epoch)
if game.is_won():
win_count += 1
print("Accuracy {} %".format(100. * win_count / nb_epoch))
'''
class Agent:
def __init__(self, model, memory=None, memory_size=1000, nb_frames=None):
assert len(model.output_shape) == 2, "Model's output shape should be (nb_samples, nb_actions)."
if memory:
self.memory = memory
else:
self.memory = ExperienceReplay(memory_size)
if not nb_frames and not model.input_shape:
raise Exception("Missing argument : nb_frames not provided")
elif not nb_frames:
nb_frames = model.input_shape[1]
elif model.input_shape[1] and nb_frames and model.input_shape[1] != nb_frames:
raise Exception("Dimension mismatch : time dimension of model should be equal to nb_frames.")
self.model = model
self.nb_frames = nb_frames
self.frames = None
@property
def memory_size(self):
return self.memory.memory_size
@memory_size.setter
def memory_size(self, value):
self.memory.memory_size = value
def reset_memory(self):
self.exp_replay.reset_memory()
def check_game_compatibility(self, game):
game_output_shape = (1, None) + game.get_frame().shape
if len(game_output_shape) != len(self.model.input_shape):
raise Exception('Dimension mismatch. Input shape of the model should be compatible with the game.')
else:
for i in range(len(self.model.input_shape)):
if self.model.input_shape[i] and game_output_shape[i] and self.model.input_shape[i] != game_output_shape[i]:
raise Exception('Dimension mismatch. Input shape of the model should be compatible with the game.')
if len(self.model.output_shape) != 2 or self.model.output_shape[1] != game.nb_actions:
raise Exception('Output shape of model should be (nb_samples, nb_actions).')
def get_game_data(self, game):
frame = game.get_frame()
if self.frames is None:
self.frames = [frame] * self.nb_frames
else:
self.frames.append(frame)
self.frames.pop(0)
return np.expand_dims(self.frames, 0)
def clear_frames(self):
self.frames = None
def action_count(self, game):
#print "game.get_action_count: ", game.get_action_count
return game.get_action_count
# SET WHICH RUNS TO PRINT OUT HERE *****************************************************************
def report_action(self, game):
return ((self.action_count(game) % self.report_freq) == 0) # and ((self.action_count(game) % self.report_freq) < 20) #% 10000) == 0 #
def train(self, game, nb_epoch=1000, batch_size=50, gamma=0.9, epsilon=[1., .1], epsilon_rate=0.5, reset_memory=False, id=""):
txt_store_path = "./txtstore/run_1000e_b50_15r_reg_lr1/junk/"
printing = False
record_weights = False
self.max_moves = game.get_max_moves()
self.report_freq = self.max_moves #50
'''fo_A = open(txt_store_path + "A.txt", "rw+")
fo_G = open(txt_store_path + "G.txt", "rw+")
fo_Gb = open(txt_store_path + "Gb.txt", "rw+")
fo_I = open(txt_store_path + "I.txt", "rw+")
fo_Q = open(txt_store_path + "Q.txt", "rw+")
fo_R = open(txt_store_path + "R.txt", "rw+")
fo_S = open(txt_store_path + "S.txt", "rw+")
fo_T = open(txt_store_path + "T.txt", "rw+")
fo_W = open(txt_store_path + "W.txt", "rw+")
fo_Wb = open(txt_store_path + "Wb.txt", "rw+")'''
self.check_game_compatibility(game)
if type(epsilon) in {tuple, list}:
delta = ((epsilon[0] - epsilon[1]) / (nb_epoch * epsilon_rate))
final_epsilon = epsilon[1]
epsilon = epsilon[0]
else:
final_epsilon = epsilon
model = self.model
nb_actions = model.output_shape[-1]
win_count = 0
scores = np.zeros((nb_epoch,self.max_moves/self.report_freq))
losses = np.zeros((nb_epoch,self.max_moves/self.report_freq))
for epoch in range(nb_epoch):
#ipdb.set_trace(context=9) # TRACING HERE *********************************************
loss = 0.
game.reset()
self.clear_frames()
if reset_memory:
self.reset_memory()
game_over = False
S = self.get_game_data(game)
no_last_S = True
plot_showing = False
while not game_over:
if np.random.random() < epsilon:
a = int(np.random.randint(game.nb_actions))
#if (self.action_count(game) % 100000) == 0:
'''if self.report_action(game):
if printing:
print "random",
q = model.predict(S)'''
q = model.predict(S)
expected_action = (a == int(np.argmax(q[0])))
else:
expected_action = True
q = model.predict(S)
#print q.shape
#print q[0]
# ************************************** CATCHING NANS
'''if (q[0,0] != q[0,0]):
ipdb.set_trace(context=9) # TRACING HERE *********************************************
'''
a = int(np.argmax(q[0]))
#if (self.action_count(game) % 100000) == 0:
prob = epsilon/game.nb_actions
if expected_action:
prob = 1 - epsilon + prob
game.play(a, self.report_action(game))
r = game.get_score()
#ipdb.set_trace(context=9) # TRACING HERE *********************************************
# PRINTING S HERE ******************************************************************
''' if plot_showing:
plt.clf()
plt.imshow(np.reshape(S,(6,6)))
plt.draw()
plt.show(block=False)
plot_showing = True
print "hi" '''
# PRINTING S HERE ******************************************************************
S_prime = self.get_game_data(game)
'''if self.report_action(game):
if printing:
print "S: ", S
#if no_last_S:
# last_S = S
# no_last_S = False
#else:
# print "dS:", S - last_S
# print " ==> Q(lS):", model.predict(last_S)
#print
print " ==> Q(S): ", q, " ==> A: ", a, " ==> R: %f" % r
#print " ==> Q(S'):", model.predict(S_prime)
#print
fo_S.seek(0,2)
np.savetxt(fo_S, S[0], fmt='%4.4f') #
fo_Q.seek(0,2)
np.savetxt(fo_Q, q, fmt='%4.4f') #
fo_A.seek(0,2)
fo_A.write(str(a)+"\n") #savetxt(fo, S[0], fmt='%4.4f') #
fo_R.seek(0,2)
fo_R.write(str(r)+"\n")
'''
#ipdb.set_trace(context=9) # TRACING HERE *********************************************
#last_S = S
game_over = game.is_over()
transition = [S, a, r, S_prime, game_over, prob]
self.memory.remember(*transition)
S = S_prime
batch = self.memory.get_batch(model=model, batch_size=batch_size, gamma=gamma, ruql=True) #, print_it=False) #self.report_action(game))
if batch:
inputs, targets, probs = batch
#print("model.total_loss: ", model.total_loss)
'''if record_weights:
weights_pre = model.get_weights() # GOT WEIGHTS *************************
#print "weights_pre"
#print weights_pre
if self.report_action(game):
fo_W.seek(0,2)
np.savetxt(fo_W, weights_pre[0], fmt='%4.4f') #
fo_W.write("\n")
fo_Wb.seek(0,2)
np.savetxt(fo_Wb, weights_pre[1], fmt='%4.4f') #
fo_Wb.write("\n")'''
#output = model.train_on_batch(inputs, targets)
#loss += float(output[0]) #model.train_on_batch(inputs, targets))
'''print "myAgent"
print 'inputs: ', type(inputs), "; ", inputs.shape
print 'targets: ', type(targets), "; ", targets.shape
print 'probs: ', type(probs), "; ", probs.shape'''
loss += float(model.train_on_batch(inputs, targets, probs=probs))
#if self.report_action(game):
# #print output
# #fo_G.seek(0,2)
# #np.savetxt(fo_G, output[1], fmt='%4.4f') #
# #fo_G.write("\n")
# #fo_Gb.seek(0,2)
# #np.savetxt(fo_Gb, output[2], fmt='%4.4f') #
# #fo_Gb.write("\n")
#weights_post = model.get_weights() # GOT WEIGHTS ********************************
#print "weights_post"
#print weights_post
#ipdb.set_trace() # TRACING HERE *********************************************
#print("action_count PRE: ", action_count)
if self.report_action(game):
action_count = self.action_count(game)
#print("action_count/self.report_freq: ", action_count/self.report_freq)
#print("action_count: ", action_count)
#print("self.report_freq: ", self.report_freq)
#print("scores so far: ", scores)
#print("scores.shape: ", scores.shape)'''
while (action_count/self.report_freq > scores.shape[1]):
scores = np.append(scores, np.zeros((nb_epoch,1)), 1)
losses = np.append(losses, np.zeros((nb_epoch,1)), 1)
scores[epoch, action_count/self.report_freq-1] = game.get_total_score()
losses[epoch, action_count/self.report_freq-1] = loss
#print ("running a batch (of %d): 1: %d; 2: %d" % (len(batch), batch[0].size, \
# batch[1].size))
#print "memory size: ", self.memory_size
#print "using memory\n", inputs, "; tgt: ", targets
#fo_I.seek(0,2)
#np.savetxt(fo_I, inputs[0], fmt='%4.4f') #
#fo_T.seek(0,2)
#np.savetxt(fo_T, targets, fmt='%4.4f') #
#fo_T.write("\n")
if game.is_won():
win_count += 1
if epsilon > final_epsilon:
epsilon -= delta
if (epoch % 50) == 0:
print("Epoch {:03d}/{:03d} | Loss {:.4f} | Epsilon {:.2f} | Win count {}".format(epoch + 1, nb_epoch, loss, epsilon, win_count))
pickle.dump(scores, open(txt_store_path + "score" + id + ".p", "wb" ) )
pickle.dump(losses, open(txt_store_path + "loss" + id + ".p", "wb" ) )
'''
fo_A.close()
fo_G.close()
fo_Gb.close()
fo_I.close()
fo_Q.close()
fo_R.close()
fo_S.close()
fo_T.close()
fo_W.close()
fo_Wb.close()'''
average_taken_over = 10
last_col = self.max_moves/self.report_freq -1
fo_log = open("log.txt", "rw+")
fo_log.seek(0,2)
average_score = np.mean(scores[-average_taken_over:nb_epoch, last_col])
average_error = np.mean(losses[-average_taken_over:nb_epoch, last_col])
fo_log.write("\n{:20}|{:^12}|{:^10}|{:^10}|{:^6}|{:^12}|{:^12}|{:^12}{:^6}{:^6}|{:^10}|{:^20}|{:^10}|{:^6}".format(" ", "game moves", "avg score", "error", "WC", "epochs", "batch size", "epsiln frm", ".. to", ".. by", "lr", "desciption", "timer", "reg"))
fo_log.write("\n{:<20}|{:^12d}|{:^10.2f}|{:^10.2f}|{:^6d}|".format(time.strftime("%d/%m/%Y %H:%M"), self.max_moves, \
average_score, average_error, win_count)) #average_taken_over,
fo_log.close()
def play(self, game, nb_epoch=1, epsilon=0., visualize=False):
self.check_game_compatibility(game)
model = self.model
win_count = 0
frames = []
for epoch in range(nb_epoch):
game.reset()
self.clear_frames()
S = self.get_game_data(game)
if visualize:
frames.append(game.draw())
game_over = False
while not game_over:
if np.random.rand() < epsilon:
print("random")
action = int(np.random.randint(0, game.nb_actions))
else:
q = model.predict(S)
action = int(np.argmax(q[0]))
game.play(action)
S = self.get_game_data(game)
if visualize:
frames.append(game.draw())
game_over = game.is_over()
if game.is_won():
win_count += 1
print("Accuracy {} %".format(100. * win_count / nb_epoch))
if visualize:
if 'images' not in os.listdir('.'):
os.mkdir('images')
for i in range(len(frames)):
plt.imshow(frames[i], interpolation='none')
plt.savefig("images/" + game.name + str(i) + ".png")
class Agent:
def __init__(self, model, memory=None, memory_size=1000, nb_frames=None):
assert len(
model.output_shape
) == 2, "Model's output shape should be (nb_samples, nb_actions)."
if memory:
self.memory = memory
else:
self.memory = ExperienceReplay(memory_size)
if not nb_frames and not model.input_shape:
raise Exception("Missing argument : nb_frames not provided")
elif not nb_frames:
nb_frames = model.input_shape[1]
elif model.input_shape[
1] and nb_frames and model.input_shape[1] != nb_frames:
raise Exception(
"Dimension mismatch : time dimension of model should be equal to nb_frames."
)
self.model = model
self.nb_frames = nb_frames
self.frames = None
@property
def memory_size(self):
return self.memory.memory_size
@memory_size.setter
def memory_size(self, value):
self.memory.memory_size = value
def reset_memory(self):
self.exp_replay.reset_memory()
def check_game_compatibility(self, game):
game_output_shape = (1, None) + game.get_frame().shape
if len(game_output_shape) != len(self.model.input_shape):
raise Exception(
'Dimension mismatch. Input shape of the model should be compatible with the game.'
)
else:
for i in range(len(self.model.input_shape)):
if self.model.input_shape[i] and game_output_shape[
i] and self.model.input_shape[i] != game_output_shape[
i]:
raise Exception(
'Dimension mismatch. Input shape of the model should be compatible with the game.'
)
if len(self.model.output_shape
) != 2 or self.model.output_shape[1] != game.nb_actions:
raise Exception(
'Output shape of model should be (nb_samples, nb_actions).')
def get_game_data(self, game):
frame = game.get_frame()
if self.frames is None:
self.frames = [frame] * self.nb_frames
else:
self.frames.append(frame)
self.frames.pop(0)
return np.expand_dims(self.frames, 0)
def clear_frames(self):
self.frames = None
def train(self,
game,
nb_epoch=1000,
batch_size=50,
gamma=0.9,
epsilon=[1., .1],
epsilon_rate=0.5,
reset_memory=False):
self.check_game_compatibility(game)
if type(epsilon) in {tuple, list}:
delta = ((epsilon[0] - epsilon[1]) / (nb_epoch * epsilon_rate))
final_epsilon = epsilon[1]
epsilon = epsilon[0]
else:
final_epsilon = epsilon
model = self.model
nb_actions = model.output_shape[-1]
win_count = 0
for epoch in range(nb_epoch):
loss = 0.
game.reset()
self.clear_frames()
if reset_memory:
self.reset_memory()
game_over = False
S = self.get_game_data(game)
while not game_over:
if np.random.random() < epsilon:
a = int(np.random.randint(game.nb_actions))
else:
q = model.predict(S)
a = int(np.argmax(q[0]))
game.play(a)
r = game.get_score()
S_prime = self.get_game_data(game)
game_over = game.is_over()
transition = [S, a, r, S_prime, game_over]
self.memory.remember(*transition)
S = S_prime
inputs, targets = self.memory.get_batch(model=model,
batch_size=batch_size,
gamma=gamma)
loss += model.train_on_batch(inputs, targets)[0]
if game.is_won():
win_count += 1
if epsilon > final_epsilon:
epsilon -= delta
print(
"Epoch {:03d}/{:03d} | Loss {:.4f} | Epsilon {:.2f} | Win count {}"
.format(epoch + 1, nb_epoch, loss, epsilon, win_count))
def play(self, game, nb_epoch=10, epsilon=0., visualize=True):
self.check_game_compatibility(game)
model = self.model
win_count = 0
frames = []
for epoch in range(nb_epoch):
game.reset()
self.clear_frames()
S = self.get_game_data(game)
if visualize:
frames.append(game.draw())
game_over = False
while not game_over:
if np.random.rand() < epsilon:
print("random")
action = int(np.random.randint(0, game.nb_actions))
else:
q = model.predict(S)
action = int(np.argmax(q[0]))
game.play(action)
S = self.get_game_data(game)
if visualize:
frames.append(game.draw())
game_over = game.is_over()
if game.is_won():
win_count += 1
print("Accuracy {} %".format(100. * win_count / nb_epoch))
if visualize:
if 'images' not in os.listdir('.'):
os.mkdir('images')
for i in range(len(frames)):
plt.imshow(frames[i], interpolation='none')
plt.savefig("images/" + game.name + str(i) + ".png")
class Agent:
def __init__(self, game, mode=SIMPLE, nb_epoch=10000, memory_size=1000, batch_size=50, nb_frames=4, epsilon=1., discount=.9, learning_rate=.1, model=None):
self.game = game
self.mode = mode
self.target_model = None
self.rows, self.columns = game.field_shape()
self.nb_epoch = nb_epoch
self.nb_frames = nb_frames
self.nb_actions = game.nb_actions()
if mode == TEST:
print('Training Mode: Loading model...')
self.model = load_model(model)
elif mode == SIMPLE:
print('Using Plain DQN: Building model...')
self.model = self.build_model()
elif mode == DOUBLE:
print('Using Double DQN: Building primary and target model...')
self.model = self.build_model()
self.target_model = self.build_model()
self.update_target_model()
# Trades off the importance of sooner versus later rewards.
# A factor of 0 means it rather prefers immediate rewards
# and it will mostly consider current rewards. A factor of 1
# will make it strive for a long-term high reward.
self.discount = discount
# The learning rate or step size determines to what extent the newly
# acquired information will override the old information. A factor
# of 0 will make the agent not learn anything, while a factor of 1
# would make the agent consider only the most recent information
self.learning_rate = learning_rate
# Use epsilon-greedy exploration as our policy.
# Epsilon determines the probability for choosing random actions.
# This factor will decrease linear by the number of epoches. So we choose
# a random action by the probability 'eps'. Without this policy the network
# is greedy and it will it settles with the first effective strategy it finds.
# Hence, we introduce certain randomness.
# Epislon reaches its minimum at 1/2 of the games
epsilon_end = self.nb_epoch - (self.nb_epoch / 2)
self.policy = EpsGreedyPolicy(self.model, epsilon_end, self.nb_actions, epsilon, .1)
# Create new experience replay memory. Without this optimization
# the training takes extremely long even on a GPU and most
# importantly the approximation of Q-values using non-linear
# functions, that is used for our NN, is not very stable.
self.memory = ExperienceReplay(self.model, self.target_model, self.nb_actions, memory_size, batch_size, self.discount, self.learning_rate)
self.frames = None
def build_model(self):
model = Sequential()
model.add(Conv2D(32, (2, 2), activation='relu', input_shape=(self.nb_frames, self.rows, self.columns), data_format="channels_first"))
model.add(Conv2D(64, (2, 2), activation='relu'))
model.add(Conv2D(64, (3, 3), activation='relu'))
model.add(Flatten())
model.add(Dropout(0.1))
model.add(Dense(512, activation='relu'))
model.add(Dense(self.nb_actions))
model.compile(Adam(), 'MSE')
return model
def update_target_model(self):
self.target_model.set_weights(self.model.get_weights())
def get_frames(self):
frame = self.game.get_state()
if self.frames is None:
self.frames = [frame] * self.nb_frames
else:
self.frames.append(frame)
self.frames.pop(0)
# Expand frames to match the input shape for the CNN (4D)
# 1D = # batches
# 2D = # frames per batch
# 3D / 4D = game board
return np.expand_dims(self.frames, 0)
def clear_frames(self):
self.frames = None
def print_stats(self, data, y_label, x_label='Epoch', marker='-'):
data = np.array(data)
x, y = data.T
p = np.polyfit(x, y, 3)
fig = plt.figure()
plt.plot(x, y, marker)
plt.plot(x, np.polyval(p, x), 'r:')
plt.xlabel(x_label)
plt.ylabel(y_label)
words = y_label.split()
file_name = '_'.join(map(lambda x: x.lower(), words))
path = './plots/{name}_{size}x{size}_{timestamp}'
fig.savefig(path.format(size=self.game.grid_size, name=file_name, timestamp=int(time())))
def train(self, update_freq=10):
total_steps = 0
max_steps = self.game.grid_size**2 * 3
loops = 0
nb_wins = 0
cumulative_reward = 0
duration_buffer = []
reward_buffer = []
steps_buffer = []
wins_buffer = []
for epoch in range(self.nb_epoch):
loss = 0.
duration = 0
steps = 0
self.game.reset()
self.clear_frames()
done = False
# Observe the initial state
state_t = self.get_frames()
start_time = time()
while(not done):
# Explore or Exploit
action = self.policy.select_action(state_t, epoch)
# Act on the environment
_, reward, done, is_victory = self.game.act(action)
state_tn = self.get_frames()
cumulative_reward += reward
steps += 1
total_steps += 1
if steps == max_steps and not done:
loops += 1
done = True
# Build transition and remember it (Experience Replay)
transition = [state_t, action, reward, state_tn, done]
self.memory.remember(*transition)
state_t = state_tn
# Get batch of batch_size samples
# A batch generally approximates the distribution of the input data
# better than a single input. The larger the batch, the better the
# approximation. However, larger batches take longer to process.
batch = self.memory.get_batch()
if batch:
inputs, targets = batch
loss += float(self.model.train_on_batch(inputs, targets))
if self.game.is_victory():
nb_wins += 1
if done:
duration = utils.get_time_difference(start_time, time())
if self.mode == DOUBLE and self.target_model is not None and total_steps % (update_freq) == 0:
self.update_target_model()
current_epoch = epoch + 1
reward_buffer.append([current_epoch, cumulative_reward])
duration_buffer.append([current_epoch, duration])
steps_buffer.append([current_epoch, steps])
wins_buffer.append([current_epoch, nb_wins])
summary = 'Epoch {:03d}/{:03d} | Loss {:.4f} | Epsilon {:.2f} | Time(ms) {:3.3f} | Steps {:.2f} | Wins {} | Loops {}'
print(summary.format(current_epoch, self.nb_epoch, loss, self.policy.get_eps(), duration, steps, nb_wins, loops))
# Generate plots
self.print_stats(reward_buffer, 'Cumulative Reward')
self.print_stats(duration_buffer, 'Duration per Game')
self.print_stats(steps_buffer, 'Steps per Game')
self.print_stats(wins_buffer, 'Wins')
path = './models/model_{mode}_{size}x{size}_{epochs}_{timestamp}.h5'
mode = 'dqn' if self.mode == SIMPLE else 'ddqn'
self.model.save(path.format(mode=mode, size=self.game.grid_size, epochs=self.nb_epoch, timestamp=int(time())))
def play(self, nb_games=5, interval=.7):
nb_wins = 0
accuracy = 0
summary = '{}\n\nAccuracy {:.2f}% | Game {}/{} | Wins {}'
for epoch in range(nb_games):
self.game.reset()
self.clear_frames()
done = False
state_t = self.get_frames()
self.print_state(summary, state_t[:,-1], accuracy, epoch, nb_games, nb_wins, 0)
while(not done):
q = self.model.predict(state_t)
action = np.argmax(q[0])
_, _, done, is_victory = self.game.act(action)
state_tn = self.get_frames()
state_t = state_tn
if is_victory:
nb_wins += 1
accuracy = 100. * nb_wins / nb_games
self.print_state(summary, state_t[:,-1], accuracy, epoch, nb_games, nb_wins, interval)
def print_state(self, summary, state, accuracy, epoch, nb_games, nb_wins, interval):
utils.clear_screen()
print(summary.format(state, accuracy, epoch + 1, nb_games, nb_wins))
sleep(interval)