def run():
pygame.init()
agent = DQNAgent()
counter_games = 0
score_plot = []
counter_plot =[]
record = 0
while counter_games < 100:
game = Game(width, height)
pipe = game.pipe
player = game.player
while not game.crash:
agent.epsilon = 60 - counter_games
state_old = agent.get_state(game, player, pipe)
if randint(0, 200) < agent.epsilon:
final_move = to_categorical(randint(0, 1), num_classes=2)
else:
# predict action based on the old state
prediction = agent.model.predict(state_old.reshape((1,6)))
final_move = to_categorical(np.argmax(prediction[0]), num_classes=2)
player.Move(game,final_move)
pipe.move_pipe(160)
##check_collision(player,pipe)
check_score(pipe,game )
reward = agent.set_reward(pipe, game.crash)
state_new = agent.get_state(game, player, pipe)
agent.train_short_memory(state_old, final_move, reward, state_new, game.crash)
# store the new data into a long term memory
agent.remember(state_old, final_move, reward, state_new, game.crash)
display(game,player,pipe,record)
if game.score == 20:
game.crash = True;
agent.replay_new(agent.memory)
counter_games += 1
print('Game', counter_games, ' Score:', game.score)
score_plot.append(game.score)
counter_plot.append(counter_games)
record = get_record(game.score, record)
agent.model.save_weights('weights.hdf5')
plot_seaborn(counter_plot, score_plot)
pygame.quit()
def run():
agent = DQNAgent(size)
counter_games = 0
score_plot = []
counter_plot = []
record = 0
while counter_games < games:
# Initialize classes
game = Game(size, size)
player1 = game.player
food1 = game.food
# Perform first move
initialize_game(player1, game, food1, agent)
if display_option:
display(player1, food1, game, record)
while not game.crash:
#agent.epsilon is set to give randomness to actions
agent.epsilon = (games * 0.4) - counter_games
#get old state
state_old = agent.get_state(game, player1, food1)
#perform random actions based on agent.epsilon, or choose the action
if randint(0, games) < agent.epsilon:
final_move = to_categorical(randint(0, 2), num_classes=3)
else:
# predict action based on the old state
prediction = agent.model.predict(state_old.reshape((1, agent.size)))
final_move = to_categorical(np.argmax(prediction[0]), num_classes=3)
#perform new move and get new state
player1.do_move(final_move, player1.x, player1.y, game, food1, agent)
state_new = agent.get_state(game, player1, food1)
#set reward for the new state
reward = agent.set_reward(player1, game.crash)
#train short memory base on the new action and state
agent.train_short_memory(state_old, final_move, reward, state_new, game.crash)
# store the new data into a long term memory
agent.remember(state_old, final_move, reward, state_new, game.crash)
record = get_record(game.score, record)
if display_option:
display(player1, food1, game, record)
pygame.time.wait(speed)
agent.replay_new(agent.memory)
counter_games += 1
score_plot.append(game.score)
counter_plot.append(counter_games)
print('Game', counter_games, ' Score:', game.score, 'Last 10 Avg:', np.mean(score_plot[-10:]))
agent.model.save_weights('weights.hdf5')
plot_seaborn(counter_plot, score_plot)
def run():
pygame.init()
agent = DQNAgent()
counter_games = 0
score_plot = []
counter_plot = []
record = 0
while counter_games < 150:
# Initialize classes
game = Game(440, 440)
player1 = game.player
food1 = game.food
# Perform first move
initialize_game(player1, game, food1, agent)
if display_option:
display(player1, food1, game, record)
while not game.crash:
# agent.epsilon is set to give randomness to actions
agent.epsilon = 80 - counter_games
# get old state
state_old = agent.get_state(game, player1, food1)
# predict action based on the old state
final_move = agent.predict(state_old)
# perform new move and get new state
player1.do_move(final_move, player1.x, player1.y, game, food1,
agent)
state_new = agent.get_state(game, player1, food1)
# set treward for the new state
reward = agent.set_reward(player1, game.crash)
# train short memory base on the new action and state
agent.train_short_memory(state_old, final_move, reward, state_new,
game.crash)
# store the new data into a long term memory
agent.remember(state_old, final_move, reward, state_new,
game.crash)
record = get_record(game.score, record)
if display_option:
display(player1, food1, game, record)
pygame.time.wait(speed)
agent.replay_new(agent.memory)
counter_games += 1
print('Game', counter_games, ' Score:', game.score)
score_plot.append(game.score)
counter_plot.append(counter_games)
agent.model.save_weights('weights.hdf5')
plot_seaborn(counter_plot, score_plot)
def test():
pygame.init()
agent = DQNAgent(output_dim=3)
agent.model.load_weights('weights.hdf5')
# while counter_games < 150:
# Initialize classes
game = Game(440, 440)
player1 = game.player
field0 = game.field
# Perform first move
record = 0
initialize_game(player1, game, field0, agent)
if display_option:
display(player1, field0, game, record)
while not game.crash:
# get old state
state_old = agent.get_state(game, player1, field0)
# predict action based on the old state
prediction = agent.model.predict(state_old)
final_move = np.argmax(prediction[0])
print("move {} with prediction : {}".format(final_move, prediction))
# perform new move and get new state
player1.do_move(final_move, field0, game)
record = get_record(game.score, record)
if display_option:
pygame.time.wait(speed)
display(player1, field0, game, record)
def run():
pygame.init()
agent = DQNAgent()
counter_games = 0
score_plot = []
counter_plot = []
record = 0
while counter_games < 150:
# Initialize classes
game = Game(440, 440)
player1 = game.player
food1 = game.food
# Perform first move
initialize_game(player1, game, food1, agent)
if display_option:
display(player1, food1, game, record)
while not game.crash:
agent.epsilon = 80 - counter_games
state_old = agent.get_state(game, player1, food1)
if randint(0, 200) < agent.epsilon:
final_move = to_categorical(randint(0, 2), num_classes=3)[0]
else:
prediction = agent.model.predict(state_old.reshape((1, 11)))
final_move = to_categorical(np.argmax(prediction[0]),
num_classes=3)[0]
player1.do_move(final_move, player1.x, player1.y, game, food1,
agent)
state_new = agent.get_state(game, player1, food1)
reward = agent.set_reward(player1, game.crash)
agent.train_short_memory(state_old, final_move, reward, state_new,
game.crash)
agent.remember(state_old, final_move, reward, state_new,
game.crash)
record = get_record(game.score, record)
if display_option:
display(player1, food1, game, record)
pygame.time.wait(speed)
agent.replay_new(agent.memory)
counter_games += 1
print('Game', counter_games, ' Score:', game.score)
score_plot.append(game.score)
counter_plot.append(counter_games)
agent.model.save_weights('weights.hdf5')
plot_seaborn(counter_plot, score_plot)
def run(params):
"""
Run the DQN algorithm, based on the parameters previously set.
"""
pygame.init()
agent = DQNAgent(params)
agent = agent.to(DEVICE)
agent.optimizer = optim.Adam(agent.parameters(),
weight_decay=0,
lr=params['learning_rate'])
counter_games = 0
score_plot = []
counter_plot = []
record = 0
total_score = 0
while counter_games < params['episodes']:
for event in pygame.event.get():
if event.type == pygame.QUIT:
pygame.quit()
quit()
# Initialize classes
game = Game(440, 440)
player1 = game.player
food1 = game.food
# Perform first move
initialize_game(player1, game, food1, agent, params['batch_size'])
if params['display']:
display(player1, food1, game, record)
while not game.crash:
if not params['train']:
agent.epsilon = 0.01
else:
# agent.epsilon is set to give randomness to actions
agent.epsilon = 1 - (counter_games *
params['epsilon_decay_linear'])
# get old state
state_old = agent.get_state(game, player1, food1)
# perform random actions based on agent.epsilon, or choose the action
if random.uniform(0, 1) < agent.epsilon:
final_move = np.eye(3)[randint(0, 2)]
else:
# predict action based on the old state
with torch.no_grad():
state_old_tensor = torch.tensor(
state_old.reshape(
(1, 11)), dtype=torch.float32).to(DEVICE)
prediction = agent(state_old_tensor)
final_move = np.eye(3)[np.argmax(
prediction.detach().cpu().numpy()[0])]
# perform new move and get new state
player1.do_move(final_move, player1.x, player1.y, game, food1,
agent)
state_new = agent.get_state(game, player1, food1)
# set reward for the new state
reward = agent.set_reward(player1, game.crash)
if params['train']:
# train short memory base on the new action and state
agent.train_short_memory(state_old, final_move, reward,
state_new, game.crash)
# store the new data into a long term memory
agent.remember(state_old, final_move, reward, state_new,
game.crash)
record = get_record(game.score, record)
if params['display']:
display(player1, food1, game, record)
pygame.time.wait(params['speed'])
if params['train']:
agent.replay_new(agent.memory, params['batch_size'])
counter_games += 1
total_score += game.score
print(f'Game {counter_games} Score: {game.score}')
score_plot.append(game.score)
counter_plot.append(counter_games)
mean, stdev = get_mean_stdev(score_plot)
if params['train']:
model_weights = agent.state_dict()
torch.save(model_weights, params["weights_path"])
if params['plot_score']:
plot_seaborn(counter_plot, score_plot, params['train'])
return total_score, mean, stdev
agent.memoize(state_init1, action, reward, state_init2, game.game_over)
agent.replay_new(agent.memory)
while num_games < 500:
num_games += 1
player = []
game = Game(window_width, window_height)
player.append(
Player(game, 0.5 * game.window_width, 0.5 * game.window_height))
food = Food(game)
init(agent, game, player, food)
while game.game_over == False:
# for random moves
agent.epsilon = epsilon - num_games
old_state = agent.get_state(game, player, food)
if random.randint(0, 200) < agent.epsilon:
new_direction = to_categorical(random.randint(0, 4), num_classes=5)
else:
predict = agent.model.predict(old_state.reshape((1, 13)))
new_direction = to_categorical(np.argmax(predict[0]),
num_classes=5)
# perform move
event_handler(player, np.argmax(np.array(new_direction)))
update_screen()
new_state = agent.get_state(game, player, food)
reward = agent.set_reward(game)
class Simulation():
def __init__(self):
self.amountOfSimulations = 0
self.maxSpeed = 4
self.score = 0
self.previousScore = 0
self.highscore = 0
self.highscoreTime = 0
#self.delayCounter = 0
#self.delay = randint(5,10)
self.keepRunning = True
self.agent = DQNAgent()
self.mysystem = chrono.ChSystemNSC()
self.ground = theBattleground.theBattleground(self.mysystem)
self.createRobot(self.mysystem)
self.createApplication()
self.run()
def createApplication(self):
# Create an Irrlicht application to visualize the system
self.myapplication = chronoirr.ChIrrApp(
self.mysystem, 'PyChrono example',
chronoirr.dimension2du(1024, 768))
self.myapplication.AddTypicalSky()
self.myapplication.AddTypicalLogo()
self.myapplication.AddTypicalCamera(chronoirr.vector3df(0.6, 0.6, 0.8))
self.myapplication.AddLightWithShadow(
chronoirr.vector3df(2, 4, 2), # point
chronoirr.vector3df(0, 0, 0), # aimpoint
9, # radius (power)
1,
9, # near, far
30) # angle of FOV
# ==IMPORTANT!== Use this function for adding a ChIrrNodeAsset to all items
# in the system. These ChIrrNodeAsset assets are 'proxies' to the Irrlicht meshes.
# If you need a finer control on which item really needs a visualization proxy in
# Irrlicht, just use application.AssetBind(myitem); on a per-item basis.
self.myapplication.AssetBindAll()
# ==IMPORTANT!== Use this function for 'converting' into Irrlicht meshes the assets
# that you added to the bodies into 3D shapes, they can be visualized by Irrlicht!
self.myapplication.AssetUpdateAll()
self.myapplication.AddShadowAll()
self.myapplication.SetShowInfos(True)
def displayScore(self):
print("Score: " + str(self.score) + " Highscore: " +
str(self.highscore))
def checkIfDead(self):
if (((self.robot.mbody1.GetRot()).Q_to_Rotv()).z < 2.5):
print("tilt forward - DEAD")
self.keepRunning = False
if (((self.robot.mbody1.GetRot()).Q_to_Rotv()).z > 4):
print("tilt backward - DEAD")
self.keepRunning = False
def doMove(self, prediction):
self.checkIfDead()
print("pred is: ")
print(prediction)
print(" ")
if prediction[0][0] > prediction[0][1]:
speed = (prediction[0][0] * self.maxSpeed)
self.robot.motor_R.SetMotorFunction(chrono.ChFunction_Const(speed))
self.robot.motor_L.SetMotorFunction(chrono.ChFunction_Const(speed))
elif prediction[0][1] > prediction[0][0]:
speed = -(prediction[0][1] * self.maxSpeed)
self.robot.motor_R.SetMotorFunction(chrono.ChFunction_Const(speed))
self.robot.motor_L.SetMotorFunction(chrono.ChFunction_Const(speed))
else:
self.robot.motor_R.SetMotorFunction(chrono.ChFunction_Const(0))
self.robot.motor_L.SetMotorFunction(chrono.ChFunction_Const(0))
def createRobot(self, system):
try:
del self.robot
except:
pass
self.robot = theRobot.theRobot(system)
def restart(self):
print("restart called")
del self.myapplication
del self.mysystem
self.mysystem = chrono.ChSystemNSC()
self.ground = theBattleground.theBattleground(self.mysystem)
self.createRobot(self.mysystem)
self.createApplication()
self.mysystem.SetChTime(0)
self.keepRunning = True
if self.score > self.highscore:
self.highscore = self.score
self.highscoreTime = self.amountOfSimulations
self.previousScore = self.score
self.score = 0
def run(self):
self.myapplication.SetTimestep(0.1)
self.myapplication.SetTryRealtime(False)
while self.amountOfSimulations <= 500:
self.restart()
while (self.myapplication.GetDevice().run() and self.keepRunning):
self.agent.epsilon = 80 - self.amountOfSimulations
#get old state
state_old = self.agent.get_state(self.robot)
self.myapplication.BeginScene()
self.myapplication.DrawAll()
self.myapplication.DoStep()
self.myapplication.EndScene()
self.checkIfDead()
#perform random actions based on agent.epsilon, or choose the action
if randint(0, 200) < self.agent.epsilon:
#prediction = to_categorical(random.random(), num_classes=2)
prediction = [[random.random(), random.random()]]
print("random action")
else:
# predict action based on the old state
print("AI action")
prediction = self.agent.model.predict(
state_old.reshape((1, 2)))
#final_move = to_categorical(np.argmax(prediction[0]), num_classes=2)
#perform new move and get new state
self.doMove(prediction)
state_new = self.agent.get_state(self.robot)
#self.delay = randint(5, 10)
self.delayCounter = 0
self.displayScore()
self.score += 1
#self.delayCounter += 1
#set reward for the new state
reward = self.agent.set_reward(self.score, self.highscore,
self.previousScore,
self.amountOfSimulations,
self.highscoreTime)
#train short memory base on the new action and state
self.agent.train_short_memory(state_old, prediction, reward,
state_new, self.keepRunning)
# store the new data into a long term memory
self.agent.remember(state_old, prediction, reward, state_new,
self.keepRunning)
self.amountOfSimulations += 1
self.agent.replay_new(self.agent.memory)
self.agent.model.save_weights('weights.hdf5')
def main():
# Initialisieren aller Pygame-Module und
# Fenster erstellen (wir bekommen eine Surface, die den Bildschirm repräsentiert).
pygame.init()
agent = DQNAgent()
counter_games = 0
record = 0
while counter_games < 150:
screen = pygame.display.set_mode((800, 600))
# Titel des Fensters setzen, Mauszeiger nicht verstecken und Tastendrücke wiederholt senden.
pygame.display.set_caption("Pygame-Tutorial: Animation")
pygame.mouse.set_visible(1)
pygame.key.set_repeat(1, 30)
pygame.font.init() # you have to call this at the start,
myfont = pygame.font.SysFont('Comic Sans MS', 30)
# Clock-Objekt erstellen, das wir benötigen, um die Framerate zu begrenzen.
clock = pygame.time.Clock()
# Wir erstellen eine Tilemap.
map = Tilemap.Tilemap()
event = AutoInput.AutoInput()
# Die Schleife, und damit unser Spiel, läuft solange running == True.
running = True
max_steps_reached = False
max_steps = 100
step = 0
max_score = map.player.pos_x
max_score_evolution = []
while running and not max_steps_reached:
agent.epsilon = 80 - counter_games
#get old state
state_old = agent.get_state(map)
map.player.pos_x_old = map.player.pos_x
#perform random actions based on agent.epsilon, or choose the action
if randint(0, 200) < agent.epsilon:
final_move = to_categorical(randint(0, 2), num_classes=3)
else:
# predict action based on the old state
prediction = agent.model.predict(state_old.reshape((1, 7)))
final_move = to_categorical(np.argmax(prediction[0]),
num_classes=3)
# Framerate auf 30 Frames pro Sekunde beschränken.
# Pygame wartet, falls das Programm schneller läuft.
clock.tick(30)
# screen Surface mit Schwarz (RGB = 0, 0, 0) füllen.
screen.fill((198, 209, 255))
map.handle_input(final_move)
#continue jump animation after
if map.player.isjump:
map.player.jump()
# Die Tilemap auf die screen-Surface rendern.
map.render(screen)
textsurface = myfont.render(
"Game " + str(counter_games) + " Step " + str(step) +
" Max Score " + str(max_score), False, (0, 0, 0))
screen.blit(textsurface, (50, 50))
#Print Hindernis onto map and check if there should be a new one
if not map.isThereHindernis:
map.createNewHindernis()
map.isThereHindernis = True
map.hindernis.move()
map.hindernis.render(screen)
map.checkHindernisOnMap()
state_new = agent.get_state(map)
crash = map.collisionDetection()
#set treward for the new state
reward = agent.set_reward(map.player, crash)
#train short memory base on the new action and state
agent.train_short_memory(state_old, final_move, reward, state_new,
running)
# Inhalt von screen anzeigen
pygame.display.flip()
if map.player.pos_x > max_score:
max_score = map.player.pos_x
step += 1
if step >= max_steps:
max_steps_reached = True
max_score_evolution.append(max_score)
agent.remember(state_old, final_move, reward, state_new, running)
#record = get_record(map.player.pos_x, record)
#if display_option:
# #display(player1, food1, game, record)
# pygame.time.wait(speed)
agent.replay_new(agent.memory)
counter_games += 1
agent.model.save_weights('weights.hdf5')
sns.plot(max_score_evolution)
def run(episodes, load_weights, display_option, speed):
pygame.init()
agent = DQNAgent()
weights_filepath = os.path.join(os.getcwd(), WEIGHTS_FILENAME)
if load_weights and os.path.isfile(weights_filepath):
agent.model.load_weights(weights_filepath)
counter_games = 0
score_plot = []
counter_plot = []
record = 0
while counter_games < episodes:
for event in pygame.event.get():
if event.type == pygame.QUIT:
pygame.quit()
quit()
# Initialize classes
game = Game(440, 440)
player1 = game.player
food1 = game.food
# Perform first move
initialize_game(player1, game, food1, agent)
if display_option:
display(player1, food1, game, record)
while not game.crash:
# agent.epsilon is set to give randomness to actions
agent.epsilon = 80 - counter_games
# get old state
state_old = agent.get_state(game, player1, food1)
# perform random actions based on agent.epsilon, or choose the action
if randint(0, 200) < agent.epsilon:
final_move = to_categorical(randint(0, 2), num_classes=3)
else:
# predict action based on the old state
prediction = agent.model.predict(state_old.reshape((1, 11)))
final_move = to_categorical(np.argmax(prediction[0]),
num_classes=3)
# perform new move and get new state
player1.do_move(final_move, player1.x, player1.y, game, food1,
agent)
state_new = agent.get_state(game, player1, food1)
# set treward for the new state
reward = agent.set_reward(player1, game.crash)
# train short memory base on the new action and state
agent.train_short_memory(state_old, final_move, reward, state_new,
game.crash)
# store the new data into a long term memory
agent.remember(state_old, final_move, reward, state_new,
game.crash)
record = get_record(game.score, record)
if display_option:
display(player1, food1, game, record)
pygame.time.wait(speed)
agent.replay_new(agent.memory)
counter_games += 1
print(f'Game {counter_games} Score: {game.score}')
score_plot.append(game.score)
counter_plot.append(counter_games)
agent.model.save_weights(WEIGHTS_FILENAME)
plot_seaborn(counter_plot, score_plot)
def run(params):
"""
Run the DQN algorithm, based on the parameters previously set.
"""
pygame.init()
agent = DQNAgent(params)
agent = agent.to(DEVICE)
agent.optimizer = optim.Adam(agent.parameters(), weight_decay=0, lr=params['learning_rate'])
counter_games = 0
score_plot = []
counter_plot = []
record = 0
total_score = 0
while counter_games < params['epoch']:
for event in pygame.event.get():
if event.type == pygame.QUIT:
pygame.quit()
quit()
# Initialize classes
game = Game(440, 440)
player1 = game.player
food1 = game.food
# Pierwszy ruch
initialize_game(player1, game, food1, agent, params['batch_size'])
if params['display']:
display(player1, food1, game, record)
steps = 0 # Ruchy od ostatniej otrzymanej nagrody
while (not game.crash) and (steps < 100):
if not params['train']:
agent.epsilon = 0.01
else:
# agent.epsilon dla losowosci akcji
agent.epsilon = 1 - (counter_games * params['epsilon_decay_linear'])
# Otrzymaj stary stan
state_old = agent.get_state(game, player1, food1)
# Wykonuj losowe akcje na podstawie agent.epsilon albo na podstawie starych
if random.uniform(0, 1) < agent.epsilon:
final_move = np.eye(3)[randint(0,2)]
else:
# Predykcja na podstawie starych akcji
with torch.no_grad():
state_old_tensor = torch.tensor(state_old.reshape((1, 11)), dtype=torch.float32).to(DEVICE)
prediction = agent(state_old_tensor)
final_move = np.eye(3)[np.argmax(prediction.detach().cpu().numpy()[0])]
# Wykonaj nowy ruch i otrzymaj nowy stan
player1.do_move(final_move, player1.x, player1.y, game, food1, agent)
state_new = agent.get_state(game, player1, food1)
# Okresl nagrode dla nowego ruchu
reward = agent.set_reward(player1, game.crash)
# Kiedy zje, zeruj kroki
if reward > 0:
steps = 0
if params['train']:
# Trenuj pamiec krotka na podstawie nowych ruchow
agent.train_short_memory(state_old, final_move, reward, state_new, game.crash)
# Zapisz pamiec nowych ruchow do pamieci dlugiej
agent.remember(state_old, final_move, reward, state_new, game.crash)
record = get_record(game.score, record)
if params['display']:
display(player1, food1, game, record)
pygame.time.wait(params['speed'])
steps+=1
if params['train']:
agent.replay_new(agent.memory, params['batch_size'])
counter_games += 1
total_score += game.score
print(f'Game {counter_games} Score: {game.score}')
score_plot.append(game.score)
counter_plot.append(counter_games)
mean, stdev = get_mean_stdev(score_plot)
if params['train']:
model_weights = agent.state_dict()
torch.save(model_weights, params["weights_path"])
if params['plot_score']:
plot_seaborn(counter_plot, score_plot, params['train'])
return total_score, mean, stdev
def main():
# Initialisation du reseau
agent = DQNAgent()
taille = 30 # Nombre de cellules = taille * taille
game = Game(taille) # Initialisation du jeu
init_ihm(game) # Initialization IHM
max_iteration, cpt_iteration = 200, 0 # Nombre de parties à jouer et Compteur de parties
rep_lines_bestScore, score_plot, counter_plot = [], [], [
] # scores, numéros de parties, lignes du meilleur score
while (cpt_iteration < max_iteration):
game = Game(taille) # Initialisation du jeu
game.calculer_lignes_jouables(
) # Calculer les lignes possibles d'etres jouées
if len(game.rep_playable_lines
) > 1: # choisir une ligne parmis les lignes jouable au hasard
game.choosed_line = game.rep_playable_lines[random.randint(
0,
len(game.rep_playable_lines) - 1)]
elif len(game.rep_playable_lines) != 0:
game.choosed_line = game.rep_playable_lines[0]
while not game.crash and game.cpt_liberte < game.liberte: # Tant qu'il reste encore des lignes possibles à jouer
agent.epsilon = 120 - cpt_iteration
state_old = agent.get_state(game) # get old state
# perform random actions based on agent.epsilon, or choose the action
if randint(0, 200) < agent.epsilon:
final_move = to_categorical(randint(0, 4), num_classes=8)
else: # predict action based on the old state
prediction = agent.model.predict(state_old.reshape((1, 80)))
final_move = to_categorical(np.argmax(prediction[0]),
num_classes=8)
game.move = final_move
vect_move = [] # liste des sorties possibles
vect_squar = [[-1, -1], [-1, 0], [-1, 1], [0, 1], [1, 1], [1, 0],
[1, -1], [0, -1]]
for i in range(len(vect_squar)):
vect_move.append([0, 0, 0, 0, 0, 0, 0,
0]) # initialisation des sorties
for i in range(len(vect_squar)):
vect_move[i][i] = 1 # Matrice identité 8*8
for i in range(len(vect_move)):
if vect_move[i] == final_move.tolist(
): # à quel sortie correspond le final_move
game.x_player = int(
game.x_player + vect_squar[i][0]
) # deplacer le jouer d'après la sortie indiqué par final_move
game.y_player = int(game.y_player + vect_squar[i][1])
game.move = vect_move[i] # sauvegarder mouvememnt
game.found = False
for i in range(
len(game.rep_playable_lines) - 1
): # si une ligne jouable commence par la position du joueur
if len(game.rep_playable_lines
) != 0 and game.rep_playable_lines[i][0][
0].x == game.x_player and game.rep_playable_lines[
i][0][0].y == game.y_player:
game.found = True
game.jouer_ligne(
game.rep_playable_lines[i]
) # Jouer la ligne tmp_line=[[cellule * 5],direction]
game.rep_lines.append(
game.rep_playable_lines[i]
) # Ajouter la ligne à la liste des lignes jouées
game.calculer_lignes_jouables(
) # Recalcule des lignes jouables
if cpt_iteration == 0 or len(
game.rep_lines) > max(score_plot):
rep_lines_bestScore = game.rep_lines # Sauvegarder la liste des lignes du meilleur score
state_new = agent.get_state(game)
reward = agent.set_reward(len(game.rep_playable_lines),
game.found)
# train short memory base on the new action and state
agent.train_short_memory(state_old, final_move, reward,
state_new, game.crash)
# store the new data into a long term memory
agent.remember(state_old, final_move, reward, state_new,
game.crash)
game.cpt_liberte = 0
break
if len(game.rep_playable_lines
) == 0: # s'il reste plus de lignes jouables
game.crash = True
state_new = agent.get_state(game)
reward = agent.set_reward(0, game.found)
# train short memory base on the new action and state
agent.train_short_memory(state_old, final_move, reward,
state_new, game.crash)
# store the new data into a long term memory
agent.remember(state_old, final_move, reward, state_new,
game.crash)
elif game.cpt_liberte > game.liberte - 2 and not game.found:
game.x_player = 13 # retourr à la case de départ
game.y_player = 13
game.crash = True
game.found = False
state_new = agent.get_state(game)
reward = agent.set_reward(0, game.found)
# train short memory base on the new action and state
agent.train_short_memory(state_old, final_move, reward,
state_new, game.crash)
# store the new data into a long term memory
agent.remember(state_old, final_move, reward, state_new,
game.crash)
elif game.found == False:
game.cpt_liberte += 1
if cpt_iteration == 0 or len(game.rep_lines) > max(
score_plot): # Si c'est le meilleur score atteint
rep_lines_bestScore = game.rep_lines # Sauvegarder la liste des lignes du meilleur score
score_plot.append(len(
game.rep_lines)) # Ajouter le score à la liste des scores
cpt_iteration += 1 # augmenter le nombre de partie
print(cpt_iteration, "-Score: ", len(game.rep_lines),
" - Best score: ", max(score_plot))
counter_plot.append(cpt_iteration) # ajouter le numéro de la partie
agent.replay_new(agent.memory)
afficher_lignes_ihm(
rep_lines_bestScore, max_iteration,
max(score_plot)) # afficher dans l'ihm les lignes du meilleur score
plot_seaborn(counter_plot,
score_plot) # afficher le shéma des scores selon les parties
mainloop()
def run(display_option, speed, params):
pygame.init()
agent = DQNAgent(params)
weights_filepath = params['weights_path']
if params['load_weights']:
agent.model.load_weights(weights_filepath)
print("weights loaded")
counter_games = 0
score_plot = []
counter_plot = []
record = 0
while counter_games < params['episodes']:
logger.info("===========================")
logger.info(f"{info_string}")
time_start = time.time()
for event in pygame.event.get():
if event.type == pygame.QUIT:
pygame.quit()
quit()
# Initialize classes
game = Game(440, 440)
player1 = game.player
food1 = game.food
# Perform first move
initialize_game(player1, game, food1, agent, params['batch_size'])
if display_option == True:
display(player1, food1, game, record)
time_start_game_update = time.time()
while not game.crash:
time_start_game_update_pygame = time.time()
if not params['train']:
agent.epsilon = 0
else:
# agent.epsilon is set to give randomness to actions
if agent.epsilon <= params['min_epsilon']:
agent.epsilon = params['min_epsilon']
else:
agent.epsilon = 1 - (counter_games *
params['epsilon_decay_linear'])
# get old state
state_old, vision = agent.get_state(game, player1, food1)
# perform random actions based on agent.epsilon, or choose the action
if randint(0, 1) < agent.epsilon:
final_move = to_categorical(randint(0, 2), num_classes=3)
else:
# predict action based on the old state
prediction = agent.model.predict(
state_old.reshape((1, params['num_input_features'])))
final_move = to_categorical(np.argmax(prediction[0]),
num_classes=3)
# perform new move and get new state
player1.do_move(final_move, player1.x, player1.y, game, food1,
agent)
state_new, vision = agent.get_state(game, player1, food1)
# set reward for the new state
reward = agent.set_reward(player1, game.crash)
time_end_game_update_pygame = time.time()
time_start_game_update_train = time.time()
if params['train']:
# train short memory base on the new action and state
agent.train_short_memory(state_old, final_move, reward,
state_new, game.crash)
# store the new data into a long term memory
agent.remember(state_old, final_move, reward, state_new,
game.crash)
time_end_game_update_train = time.time()
time_start_game_update_record = time.time()
record = get_record(game.score, record)
time_end_game_update_record = time.time()
logger.debug("Pygame update step: " +
str((time_end_game_update_pygame -
time_start_game_update_pygame)))
logger.debug("Train short term update step: " +
str((time_end_game_update_train -
time_start_game_update_train)))
logger.debug("Record score step: " +
str((time_end_game_update_record -
time_start_game_update_record)))
if display_option == True:
cv2.imshow("Vision of the Snake", vision * 255.0)
# detect any kepresses
key = cv2.waitKey(1) & 0xFF
# if the `q` key was pressed, break from the loop
if key == ord("q"):
break
display(player1, food1, game, record)
pygame.time.wait(speed)
# # Pause visualisation if crash
# if display_option==True:
# cv2.imshow("Vision of the Snake", vision * 255.0)
#
# # detect any kepresses
# key = cv2.waitKey(1) & 0xFF
# # if the `q` key was pressed, break from the loop
# if key == ord("q"):
# break
# display(player1, food1, game, record)
# pygame.time.wait(5000)
time_end_game_update = time.time()
logger.info(
"Time to play one game: " +
str(round((time_end_game_update - time_start_game_update), 3)))
time_start_long_term = time.time()
if params['train']:
agent.replay_new(agent.memory, params['batch_size'])
time_end_long_term = time.time()
logger.info("Train long term update step: " +
str(round((time_end_long_term - time_start_long_term), 3)))
if agent.epsilon <= params['min_epsilon']:
agent.epsilon = params['min_epsilon']
else:
agent.epsilon = 1 - (counter_games *
params['epsilon_decay_linear'])
logger.info(f'The epsilon value is: {agent.epsilon}')
logger.debug("===========================")
counter_games += 1
logger.info(f'Game {counter_games} Score: {game.score}')
logger.info(f'The agent memory length is: {len(agent.memory)}')
score_plot.append(game.score)
counter_plot.append(counter_games)
if params['train'] and counter_games % 100 == 0:
agent.model.save_weights(params['weights_path'])
logger.info("===========SAVING THE MODEL================")
with open(params['memory_path'], 'wb') as handle:
pickle.dump(agent.memory, handle)
logger.info("End Game Loop")
time_end = time.time()
epoch_timer = round((time_end - time_start), 3)
logger.info(f"One epoch takes: {epoch_timer} seconds")
eta_prediction = round(
(params['episodes'] - counter_games) * epoch_timer / 60)
logger.info(f"Time remaining is: {eta_prediction} minutes")
if params['train']:
agent.model.save_weights(params['weights_path'])
with open(params['memory_path'], 'wb') as handle:
pickle.dump(agent.memory, handle)
params['counter_plot'] = counter_plot
params['score_plot'] = score_plot
with open(params['params_path'], 'wb') as handle:
pickle.dump(params, handle)
def train(epoch=10):
pygame.init()
agent = DQNAgent(output_dim=3)
counter_games = 0
score_plot = []
counter_plot = []
record = 0
while counter_games < epoch:
# Initialize classes
game = Game(440, 440)
player1 = game.player
field0 = game.field
# Perform first move
initialize_game(player1, game, field0, agent)
if display_option:
display(player1, field0, game, record)
game_epoch = 0
while not game.crash:
#agent.epsilon is set to give randomness to actions
agent.epsilon = 50 - game_epoch
#get old state
state_old = agent.get_state(game, player1, field0)
#perform random actions based on agent.epsilon, or choose the action
if randint(0, 100) < agent.epsilon:
final_move = randint(0, 2)
# print("random with prob {}".format(agent.epsilon))
else:
# predict action based on the old state
prediction = agent.model.predict(state_old)
final_move = np.argmax(prediction[0])
print("prediction : {}".format(prediction))
# print("move: {} to position ({}, {})".format(final_move, player1.x, player1.y))
#perform new move and get new state
player1.do_move(final_move, field0, game)
if game_epoch >= 19:
# get new state
state_new = agent.get_state(game, player1, field0)
#set treward for the new state
reward = agent.set_reward(player1, game.crash, final_move)
#train short memory base on the new action and state
agent.train_short_memory(state_old, final_move, reward,
state_new, game.crash)
# store the new data into a long term memory
if_remember = False
if game.crash:
agent.remember(state_old, final_move, reward, state_new,
game.crash)
if_remember = True
# print("remember this move with reward {}".format(reward))
elif final_move == 0 and randint(1, 20) < 20:
agent.remember(state_old, final_move, reward, state_new,
game.crash)
if_remember = True
# print("remember this move with reward {}".format(reward))
elif final_move != 0 and randint(1, 20) < 20:
agent.remember(state_old, final_move, reward, state_new,
game.crash)
if_remember = True
# print("remember this move with reward {}".format(reward))
print(
"actual move {} to ({}, {}) gets reward {} - remember {}".
format(final_move, player1.x, player1.y, reward,
if_remember))
# explore other move
if final_move == 0: # no
# 1 left
explore_moves(game, field0, agent, player1, state_old, 1,
max(0, player1.x - 1), player1.y)
# 2 right
explore_moves(game, field0, agent, player1, state_old, 2,
min(player1.x + 1, 21), player1.y)
elif final_move == 1: # left
# 0 no
explore_moves(game, field0, agent, player1, state_old, 0,
min(player1.x + 1, 21), player1.y)
# 2 right
explore_moves(game, field0, agent, player1, state_old, 2,
min(player1.x + 2, 21), player1.y)
elif final_move == 2: # right
# 0 no
explore_moves(game, field0, agent, player1, state_old, 0,
max(0, player1.x - 1), player1.y)
# 2 right
explore_moves(game, field0, agent, player1, state_old, 1,
max(0, player1.x - 2), player1.y)
record = get_record(game.score, record)
if display_option:
display(player1, field0, game, record)
pygame.time.wait(speed)
game_epoch += 1
agent.replay_new(agent.memory)
counter_games += 1
print('Game', counter_games, ' Score:', game.score)
score_plot.append(game.score)
counter_plot.append(counter_games)
if game.score >= record:
agent.model.save_weights(modelFile + '/weights.hdf5')
agent.model.save_weights(modelFile + '/weightsFinal.hdf5')
plot_seaborn(counter_plot, score_plot)
def run(display_option, speed, params):
if display_option:
pygame.init()
agent = DQNAgent(params)
counter_games = 0
score_plot = []
counter_plot = []
record = 0
while counter_games < params['episodes']:
# Initialize classes
game = Game(440, 440, display_option)
player1 = game.player
food1 = game.food
# Perform first move
initialize_game(player1, game, food1, agent, params['batch_size'])
if display_option:
display(player1, food1, game, record)
step_count = 0
raw_data = [] if params['raw_output'] is not None else None
while not game.crash:
if display_option:
for event in pygame.event.get():
if event.type == pygame.QUIT:
pygame.quit()
quit()
if not params['train']:
agent.epsilon = 0
else:
# agent.epsilon is set to give randomness to actions
agent.epsilon = 1 - (counter_games *
params['epsilon_decay_linear'])
# get old state
state_old = agent.get_state(game, player1, food1)
# perform random actions based on agent.epsilon, or choose the action
if random() < agent.epsilon or random() < (step_count - 300) / 300:
prediction = [0, 0, 0]
final_move = to_categorical(randint(0, 2), num_classes=3)
else:
# predict action based on the old state
prediction = agent.model.predict(state_old.reshape((1, 11)))[0]
final_move = to_categorical(np.argmax(prediction),
num_classes=3)
if raw_data is not None:
step_data = {
'head_x': player1.x,
'head_y': player1.y,
'food_x': food1.x_food,
'food_y': food1.y_food,
'snake_position': copy.deepcopy(player1.position),
'snake_x_change': player1.x_change,
'snake_y_change': player1.y_change,
'action': final_move.tolist()
}
# perform new move and get new state
player1.do_move(final_move, player1.x, player1.y, game, food1,
agent)
game.player.action = final_move
state_new = agent.get_state(game, player1, food1)
# set reward for the new state
reward = agent.set_reward(player1, game.crash)
if raw_data is not None:
step_data['eaten'] = player1.eaten
step_data['reward'] = reward
step_data['crash'] = game.crash
step_data['next_state'] = {
'head_x': player1.x,
'head_y': player1.y,
'food_x': food1.x_food,
'food_y': food1.y_food,
'snake_position': copy.deepcopy(player1.position),
'snake_x_change': player1.x_change,
'snake_y_change': player1.y_change,
}
raw_data.append(step_data)
if params['verbose']:
print(prediction, final_move, reward, step_count)
if params['train']:
# train short memory base on the new action and state
agent.train_short_memory(state_old, final_move, reward,
state_new, game.crash)
# store the new data into a long term memory
agent.remember(state_old, final_move, reward, state_new,
game.crash)
record = get_record(game.score, record)
if display_option:
display(player1, food1, game, record)
pygame.time.wait(speed)
step_count += 1
if params['train']:
agent.replay_new(agent.memory, params['batch_size'])
counter_games += 1
print(
f'Game {counter_games}/{step_count}/{agent.epsilon} Score: {game.score}'
)
if raw_data is not None:
fn_index = params['raw_output_index'] + counter_games
with open(os.path.join(params['raw_output'], f'{fn_index}.json'),
'w') as out:
json.dump(raw_data, out)
score_plot.append(game.score)
counter_plot.append(counter_games)
if params['train']:
agent.model.save_weights(params['weights_path'])
plot_seaborn(counter_plot, score_plot)
def run():
# Workaround for the current problem of incopability to wirk with CUDA and TF -> using CPU this way in code
import os
import tensorflow as tf
# os.environ['CUDA_VISIBLE_DEVICES'] = '-1'
os.environ['CUDA_VISIBLE_DEVICES'] = ''
if tf.test.gpu_device_name():
print('[DEBUG] GPU found')
else:
print("[DEBUG] No GPU found")
# till here Workaround
pygame.init()
agent = DQNAgent()
counter_games = 0
score_plot = []
counter_plot = []
record = 0
# while counter_games < 150:
while counter_games < 100:
# Initialize classes
game = Game(440, 440)
player1 = game.player
food1 = game.food
# Perform first move
initialize_game(player1, game, food1, agent)
if display_option:
display(player1, food1, game, record)
while not game.crash:
#agent.epsilon is set to give randomness to actions
agent.epsilon = 80 - counter_games
#get old state
state_old = agent.get_state(game, player1, food1)
#perform random actions based on agent.epsilon, or choose the action
if randint(0, 200) < agent.epsilon:
final_move = to_categorical(randint(0, 2), num_classes=3)
print(f"final_move: {final_move}")
# from here random responses for starting the learning
final_response = {"type": "endRound"}
# till here random responses for starting the learning
else:
# predict action based on the old state
prediction = agent.model.predict(state_old.reshape((1, 11)))
final_move = to_categorical(np.argmax(prediction[0]),
num_classes=3)
#perform new move and get new state
player1.do_move(final_move, player1.x, player1.y, game, food1,
agent)
state_new = agent.get_state(game, player1, food1)
#set treward for the new state
reward = agent.set_reward(player1, game.crash)
#train short memory base on the new action and state
agent.train_short_memory(state_old, final_move, reward, state_new,
game.crash)
# store the new data into a long term memory
agent.remember(state_old, final_move, reward, state_new,
game.crash)
record = get_record(game.score, record)
if display_option:
display(player1, food1, game, record)
pygame.time.wait(speed)
agent.replay_new(agent.memory)
counter_games += 1
print('Game', counter_games, ' Score:', game.score)
score_plot.append(game.score)
counter_plot.append(counter_games)
agent.model.save_weights('weights.hdf5')
plot_seaborn(counter_plot, score_plot)
def run():
counter_games = 0
score_plot = []
counter_plot = []
record = 0
agent = DQNAgent(TEST_MODE)
if TEST_MODE:
pygame.init()
pygame.font.init()
print('Start testing the trained model ...')
else:
print('Start training ...')
while counter_games < 150:
# Initialize classes
game = Game(440, 440)
player1 = game.player
food1 = game.food
# Perform first move
initialize_game(player1, game, food1, agent)
if TEST_MODE:
display(player1, food1, game, record)
while not game.crash:
#agent.epsilon is set to give randomness to actions
agent.epsilon = 80 - counter_games
#get old state
state_old = agent.get_state(game, player1, food1)
#perform random actions based on agent.epsilon, or choose the action
if randint(0, 200) < agent.epsilon and not TEST_MODE:
final_move = to_categorical(randint(0, 2), num_classes=3)
else:
# predict action based on the old state
prediction = agent.model.predict(state_old)
final_move = to_categorical(np.argmax(prediction[0]),
num_classes=3)
#perform new move and get new state
player1.do_move(final_move, player1.x, player1.y, game, food1,
agent)
state_new = agent.get_state(game, player1, food1)
#set treward for the new state
reward = agent.set_reward(player1, game.crash)
#train short memory base on the new action and state
agent.train_short_memory(state_old, final_move, reward, state_new,
game.crash)
# store the new data into a long term memory
agent.remember(state_old, final_move, reward, state_new,
game.crash)
record = get_record(game.score, record)
if TEST_MODE:
display(player1, food1, game, record)
pygame.time.wait(speed)
agent.replay_new(agent.memory)
counter_games += 1
print('Game', counter_games, ' Score:', game.score)
score_plot.append(game.score)
counter_plot.append(counter_games)
agent.model.save_weights('weights.hdf5')
plot_seaborn(counter_plot, score_plot)
def gameLoop():
game_over = False
games = 0
games_to_play = 300
highscore = 0
score_plot = []
games_plot = []
agent = DQNAgent()
while games < games_to_play: # Play a total of x games
xpos = dis_width / 2 # X Spawn point coordinate
ypos = dis_height / 2 # Y Spawn point coordinate
xdir = 0
ydir = 0
snake_List = []
length_of_snake = 1
food = random_food(snake_List)
xfood = food[0]
yfood = food[1]
while not game_over:
agent.epsilon = 80 - games
# agent.epsilon = 0
state_old = agent.get_state(xpos, ypos, xdir, ydir, snake_block,
xfood, yfood, dis_width, dis_height,
snake_List)
if randint(0, 200) < agent.epsilon:
final_move = to_categorical(randint(0, 2), num_classes=3)
else:
prediction = agent.model.predict(state_old.reshape((1, 11)))
final_move = to_categorical(np.argmax(prediction[0]),
num_classes=3)
# Do Action -------------------------------------------------------------
if np.array_equal(final_move, [1, 0, 0]):
pass
elif np.array_equal(
final_move,
[0, 1, 0]) and ydir is 0: # right - going horizontal
ydir = snake_block
xdir = 0
elif np.array_equal(
final_move,
[0, 1, 0]) and xdir is 0: # right - going vertical
xdir = snake_block
ydir = 0
elif np.array_equal(
final_move,
[0, 0, 1]) and ydir is 0: # left - going horizontal
ydir = -snake_block
xdir = 0
elif np.array_equal(
final_move,
[0, 0, 1]) and xdir is 0: # left - going vertical
xdir = -snake_block
ydir = 0
# Did Action -------------------------------------------------------------
# Update Frame after Action ----------------------------------------------
eaten = False
# If collide with border
if xpos == dis_width - snake_block and xdir > 0 or xpos == 0 and xdir < 0 or ypos == dis_height - snake_block and ydir > 0 or ypos == 0 and ydir < 0:
game_over = True
xpos += xdir
ypos += ydir
dis.fill(blue)
pygame.draw.rect(dis, green,
[xfood, yfood, snake_block, snake_block])
snake_Head = [xpos, ypos]
snake_List.append(snake_Head)
if len(snake_List) > length_of_snake:
del snake_List[0]
for x in snake_List[:-1]:
if x == snake_Head:
game_over = True
draw_snake(snake_block, snake_List)
highscore = get_highscore(highscore, length_of_snake - 1)
your_score(highscore, length_of_snake - 1)
if xpos == xfood and ypos == yfood:
eaten = True
food = random_food(snake_List)
xfood = food[0]
yfood = food[1]
length_of_snake += 1
# Updated Frame after Action ---------------------------------------------
state_new = agent.get_state(xpos, ypos, xdir, ydir, snake_block,
xfood, yfood, dis_width, dis_height,
snake_List)
reward = agent.set_reward(game_over, eaten)
agent.train_short_memory(state_old, final_move, reward, state_new,
game_over)
agent.remember(state_old, final_move, reward, state_new, game_over)
pygame.display.update()
clock.tick(snake_speed)
print("Game:", games + 1, "Score:", length_of_snake - 1, "Highscore:",
highscore)
agent.replay_new(agent.memory)
games += 1
score_plot.append(length_of_snake - 1)
games_plot.append(games)
game_over = False
agent.model.save_weights('weights10x10V3.hdf5')
pygame.quit()
# Plot stats of game:
sns.set(color_codes=True)
ax = sns.regplot(x=games_plot, y=score_plot)
ax.set(xlabel='games', ylabel='score')
plt.show()
quit()
def run(params):
pygame.init()
agent = DQNAgent(params)
print_info(params)
weights_filepath = params['weights_path']
if params['load_weights']:
agent.model.load_weights(weights_filepath)
print("Weights Loaded")
else:
print("Training From Scratch...")
counter_games = 0
score_plot = []
counter_plot = []
record = 0
while counter_games < params['episodes']:
for event in pygame.event.get():
if event.type == pygame.QUIT:
pygame.quit()
quit()
# Class Objects
game = Game(params)
player1 = game.player
food1 = game.food
total_reward = -150
total_reward2 = 0
# First Move
initialize_game(player1, game, food1, agent, params['batch_size'],
counter_games)
if params['display']:
display(player1, food1, game, record, counter_games, total_reward,
params)
while not game.crash:
if not params['train']:
agent.epsilon = 0
else:
# agent.epsilon is set to give randomness to actions
agent.epsilon = 1 - (
(counter_games) * params['epsilon_decay_linear'])
# old State
state_old = agent.get_state(game, player1, food1, params)
# Random Actions or Choose
if randint(0, 1) < agent.epsilon and not params['load_weights']:
final_move = to_categorical(randint(0, 2), num_classes=3)
# Prediction
else:
prediction = agent.model.predict(state_old.reshape((1, 20)))
final_move = to_categorical(np.argmax(prediction[0]),
num_classes=3)
player1.do_move(final_move, player1.x, player1.y, game, food1,
agent)
state_new = agent.get_state(game, player1, food1, params)
reward = agent.set_reward(player1, game.crash, food1,
counter_games, final_move)
total_reward += reward
total_reward2 += reward
total_reward = round(total_reward, 2)
total_reward2 = round(total_reward2, 2)
if params['train']:
# train short memory base on the new action and state
agent.train_short_memory(state_old, final_move, reward,
state_new, game.crash)
# store the new data into a long term memory
agent.remember(state_old, final_move, reward, state_new,
game.crash)
record = get_record(game.score, record)
if params['display']:
display(player1, food1, game, record, counter_games,
total_reward, params)
if params['train']:
agent.replay_new(agent.memory, params['batch_size'])
counter_games += 1
print(
f'Game {counter_games} Score: {game.score} Reward: {total_reward2}'
)
score_plot.append(game.score)
counter_plot.append(counter_games)
if counter_games % params['cf'] == 0:
n = int(counter_games / params['cf'])
agent.model.save_weights(params['cp'] + str(n) + ".hdf5")
print("Checkpoint Saved...")
plot_seaborn(counter_plot, score_plot)
pygame.quit()
quit()
def run(display_option, speed, params):
pygame.init()
agent = DQNAgent(params)
weights_filepath = params['weights_path']
if params['load_weights']:
agent.model.load_weights(weights_filepath)
print("weights loaded")
counter_games = 0
score_plot = []
counter_plot = []
record = 0
while counter_games < params['episodes']:
for event in pygame.event.get():
if event.type == pygame.QUIT:
pygame.quit()
quit()
# Initialize classes
game = Game(440, 440)
player1 = game.player
food1 = game.food
# Perform first move
initialize_game(player1, game, food1, agent, params['batch_size'])
if display_option:
display(player1, food1, game, record)
while not game.crash:
if not params['train']:
agent.epsilon = 0
else:
# agent.epsilon is set to give randomness to actions
agent.epsilon = 1 - (counter_games *
params['epsilon_decay_linear'])
# get old state
state_old = agent.get_state(game, player1, food1)
# perform random actions based on agent.epsilon, or choose the action
if randint(0, 1) < agent.epsilon:
final_move = to_categorical(randint(0, 2), num_classes=3)
else:
# predict action based on the old state
prediction = agent.model.predict(state_old.reshape((1, 11)))
final_move = to_categorical(np.argmax(prediction[0]),
num_classes=3)
# perform new move and get new state
player1.do_move(final_move, player1.x, player1.y, game, food1,
agent)
state_new = agent.get_state(game, player1, food1)
# set reward for the new state
reward = agent.set_reward(player1, game.crash)
if params['train']:
# train short memory base on the new action and state
agent.train_short_memory(state_old, final_move, reward,
state_new, game.crash)
# store the new data into a long term memory
agent.remember(state_old, final_move, reward, state_new,
game.crash)
record = get_record(game.score, record)
if display_option:
display(player1, food1, game, record)
pygame.time.wait(speed)
if params['train']:
agent.replay_new(agent.memory, params['batch_size'])
counter_games += 1
print(f'Game {counter_games} Score: {game.score}')
score_plot.append(game.score)
counter_plot.append(counter_games)
if params['train']:
agent.model.save_weights(params['weights_path'])
plot_seaborn(counter_plot, score_plot)
def run(display_option, speed, params):
pygame.init()
agent1 = DQNAgent(params, 1)
agent2 = DQNAgent(params, 2)
weights_filepath1 = params['weights_path1']
weights_filepath2 = params['weights_path2']
counter_games = 0
record1 = 0
record2 = 0
while counter_games < params['episodes']:
if params['load_weights']:
agent1.model.load_weights(weights_filepath1)
agent2.model.load_weights(weights_filepath2)
print("weights loaded")
for event in pygame.event.get():
if event.type == pygame.QUIT:
pygame.quit()
quit()
# Initialize classes
game = Game(900, 600)
player1 = Player(game, 300, 300)
player2 = Player(game, 600, 300)
# Perform first move
initialize_game(player1, player2, game, agent1, agent2,
params['batch_size'])
if display_option:
display(player1, player2, game, record1, record2)
while not game.crash:
if not params['train']:
agent1.epsilon = 0
agent2.epsilon = 0
else:
# agent.epsilon is set to give randomness to actions
agent1.epsilon = 1 - (counter_games *
params['epsilon_decay_linear'])
agent2.epsilon = 1 - (counter_games *
params['epsilon_decay_linear'])
# get old state
state_old1 = agent1.get_state(game, player1, player2)
state_old2 = agent2.get_state(game, player2, player1)
# perform random actions based on agent.epsilon, or choose the action
if randint(0, 1) < agent1.epsilon:
final_move1 = to_categorical(randint(0, 2), num_classes=3)
else:
# predict action based on the old state
prediction1 = agent1.model.predict(state_old1.reshape((1, 7)))
final_move1 = to_categorical(np.argmax(prediction1[0]),
num_classes=3)
if randint(0, 1) < agent2.epsilon:
final_move2 = to_categorical(randint(0, 2), num_classes=3)
else:
# predict action based on the old state
prediction2 = agent2.model.predict(state_old2.reshape((1, 7)))
final_move2 = to_categorical(np.argmax(prediction2[0]),
num_classes=3)
# perform new move and get new state
player1.do_move(final_move1, player1.x, player1.y, game, agent1,
player2)
player2.do_move(final_move2, player2.x, player2.y, game, agent2,
player1)
state_new1 = agent1.get_state(game, player1, player2)
state_new2 = agent2.get_state(game, player2, player1)
# set reward for the new state
reward1 = agent1.set_reward(player1, player1.crash)
reward2 = agent2.set_reward(player2, player2.crash)
if params['train']:
# train short memory base on the new action and state
agent1.train_short_memory(state_old1, final_move1, reward1,
state_new1, player1.crash)
# store the new data into a long term memory
agent1.remember(state_old1, final_move1, reward1, state_new1,
player1.crash)
agent2.train_short_memory(state_old2, final_move2, reward2,
state_new2, player2.crash)
# store the new data into a long term memory
agent2.remember(state_old2, final_move2, reward2, state_new2,
player2.crash)
game.score1 += reward1
game.score2 += reward2
record1 = get_record(game.score1, record1)
record2 = get_record(game.score2, record2)
if display_option:
display(player1, player2, game, record1, record2)
pygame.time.wait(speed)
if player1.crash and player2.crash:
game.crash = True
counter_games += 1
game.crash = False
player1.crash = False
player2.crash = False
print("score1: " + str(game.score1) + "/n")
print("score2: " + str(game.score2) + "/n")
if params['train'] and counter_games % 10 == 0:
agent1.model.save_weights(params['weights_path1'])
agent2.model.save_weights(params['weights_path2'])
print("weights saved")
if params['train']:
agent1.model.save_weights(params['weights_path1'])
agent2.model.save_weights(params['weights_path2'])
def main():
# Initialisation du reseau
agent = DQNAgent()
taille = 30 # Nombre de cellules = taille * taille
game = Game(taille) # Initialisation du jeu
init_ihm(game) # Initialization IHM
max_iteration, cpt_iteration = 2, 0 # Nombre de parties à jouer et Compteur de parties
rep_lines_bestScore, score_plot, counter_plot = [], [], [
] # scores, numéros de parties, lignes du meilleur score
while (cpt_iteration < max_iteration):
game = Game(taille) # Initialisation du jeu
game.calculer_lignes_jouables(
) # Calculer les lignes possibles d'etres jouées
while not game.crash: # Tant qu'il reste encore des lignes possibles à jouer
agent.epsilon = 80 - cpt_iteration
state_old = agent.get_state(game) # get old state
# perform random actions based on agent.epsilon, or choose the action
if randint(0, 200) < agent.epsilon:
final_move = to_categorical(randint(0, 2), num_classes=3)
else: # predict action based on the old state
prediction = agent.model.predict(state_old.reshape((1, 13)))
final_move = to_categorical(np.argmax(prediction[0]),
num_classes=3)
if np.array_equal(final_move, [1, 0, 0]):
tmp_line = game.rep_playable_lines[random.randint(
0, int((len(game.rep_playable_lines) - 1) / 3))]
elif np.array_equal(final_move, [0, 1, 0]):
tmp_line = game.rep_playable_lines[random.randint(
int((len(game.rep_playable_lines) - 1) / 3),
int((len(game.rep_playable_lines) - 1) * 2 / 3))]
elif np.array_equal(final_move, [0, 0, 1]):
tmp_line = game.rep_playable_lines[random.randint(
int((len(game.rep_playable_lines) - 1) * 2 / 3),
int(len(game.rep_playable_lines) - 1))]
game.jouer_ligne(
tmp_line) # Jouer la ligne tmp_line=[[cellule * 5],direction]
game.rep_lines.append(
tmp_line) # Ajouter la ligne à la liste des lignes jouées
game.calculer_lignes_jouables() # Recalcule des lignes jouables
best = 0
if cpt_iteration == 0 or len(game.rep_lines) > max(score_plot):
rep_lines_bestScore = game.rep_lines # Sauvegarder la liste des lignes du meilleur score
best = 1
if (len(game.rep_playable_lines) == 0
): # s'il reste plus de ligne jouable
game.crash = True
else:
state_new = agent.get_state(game)
reward = agent.set_reward(game.crash, best)
# train short memory base on the new action and state
agent.train_short_memory(state_old, final_move, reward,
state_new, game.crash)
# store the new data into a long term memory
agent.remember(state_old, final_move, reward, state_new,
game.crash)
if cpt_iteration == 0 or len(game.rep_lines) > max(
score_plot): # Si c'est le meilleur score atteint
rep_lines_bestScore = game.rep_lines # Sauvegarder la liste des lignes du meilleur score
score_plot.append(len(
game.rep_lines)) # Ajouter le score à la liste des scores
cpt_iteration += 1
print(cpt_iteration, "-Score: ", len(game.rep_lines),
" - Best score: ", max(score_plot))
counter_plot.append(cpt_iteration) # ajouter le numéro de la partie
agent.replay_new(agent.memory)
agent.model.save_weights('weights.hdf5')
afficher_lignes_ihm(rep_lines_bestScore, max_iteration, max(score_plot))
plot_seaborn(counter_plot, score_plot)
mainloop()
def run():
agent1 = DQNAgent()
agent2 = DQNAgent()
counter_games = 0
game_engine = Game()
while counter_games < GAMES_COUNT:
board = game_engine.get_init_board()
start = timer()
#set player for agents
agent1.player = randint(1, 2)
agent2.player = 1 if agent1.player == 2 else 2
while not game_engine.is_finished(board):
#agent.epsilon is set to give randomness to actions
agent1.epsilon = EPSILON - counter_games
agent2.epsilon = EPSILON - counter_games
#get state
state1 = agent1.get_state(board)
state2 = agent2.get_state(board)
#perform random actions based on agent.epsilon, or choose the action
if randint(0, EPSILON * RANDOM_MOVES_PROPORTION) < agent1.epsilon:
final_move1 = random.choice(agent1.possible_moves(state1))
else:
# predict action based on the state
prediction = agent1.model.predict(state1)
final_move1 = np.argmax(prediction[0])
if randint(0, EPSILON * RANDOM_MOVES_PROPORTION) < agent2.epsilon:
final_move2 = random.choice(agent2.possible_moves(state2))
else:
# predict action based on the state
prediction = agent2.model.predict(state2)
final_move2 = np.argmax(prediction[0])
#perform new move and get new state
board_new, changed_dir1 = game_engine.make_move(
board, final_move1, agent1.player)
board_new, changed_dir2 = game_engine.make_move(
board_new, final_move2, agent2.player)
board_new = game_engine.send_move(board_new)
state_new1 = agent1.get_state(board_new)
state_new2 = agent2.get_state(board_new)
#set reward for the new state
reward1 = agent1.get_reward(game_engine, board, board_new,
changed_dir1)
reward2 = agent2.get_reward(game_engine, board, board_new,
changed_dir2)
#train short memory base on the new action and state
game_is_finished = game_engine.is_finished(board_new)
agent1.train_short_memory(state1, final_move1, reward1, state_new1,
game_is_finished)
agent2.train_short_memory(state2, final_move2, reward2, state_new2,
game_is_finished)
# store the new data into a long term memory
agent1.remember(state1, final_move1, reward1, state_new1,
game_is_finished)
agent2.remember(state2, final_move2, reward2, state_new2,
game_is_finished)
board = board_new
if game_is_finished:
if game_engine.is_win(board, agent1.player):
agent1.wins_count += 1
if game_engine.is_win(board, agent2.player):
agent2.wins_count += 1
print('.', end='', flush=True)
agent1.replay_new()
agent2.replay_new()
counter_games += 1
print('Finished')
print('Game', counter_games)
print('Time', timer() - start)
print('Turns', board["turn"])
print('Agent 1 wins', agent1.wins_count)
print('Agent 2 wins', agent2.wins_count)
# save trained model
if agent1.wins_count > agent2.wins_count:
agent1.model.save_weights('weights.hdf5')
else:
agent2.model.save_weights('weights.hdf5')
def main():
# Initialisation du reseau
pygame.init()
agent = DQNAgent()
score_plot = []
counter_plot = []
record = 0
taille = 30 # Nombre de cellules = taille * taille
init_ihm() # Initialization IHM
max_iteration = 10 # Nombre de parties
cpt_iteration = 0 # Compteur de parties
best_score = 0 # Meilleur score
rep_lines_bestScore = [] # Liste des lignes du meilleur score
while (cpt_iteration < max_iteration):
game = Game(taille) # Initialisation du jeu
nbr_lignes_crees = 0 # Nombre de lignes crées
game.rep_lines.clear() # Mettre la liste des lignes crées à 0
game.calculer_lignes_jouables(
) # Calculer les lignes possibles d'etres jouées
while not game.crash: # Tant qu'il reste encore des lignes possibles à jouer
agent.epsilon = 80 - cpt_iteration
state_old = agent.get_state(game) # get old state
#perform random actions based on agent.epsilon, or choose the action
if randint(0, 200) < agent.epsilon:
final_move = to_categorical(randint(0, 2), num_classes=3)
else: # predict action based on the old state
prediction = agent.model.predict(state_old.reshape((1, 13)))
final_move = to_categorical(np.argmax(prediction[0]),
num_classes=3)
if np.array_equal(final_move, [1, 0, 0]):
tmp_line = game.rep_playable_lines[random.randint(
0, int((len(game.rep_playable_lines) - 1) / 3))]
elif np.array_equal(final_move, [0, 1, 0]):
tmp_line = game.rep_playable_lines[random.randint(
int((len(game.rep_playable_lines) - 1) / 3),
int((len(game.rep_playable_lines) - 1) * 2 / 3))]
elif np.array_equal(final_move, [0, 0, 1]):
tmp_line = game.rep_playable_lines[random.randint(
int((len(game.rep_playable_lines) - 1) * 2 / 3),
int(len(game.rep_playable_lines) - 1))]
game.jouer_ligne(
tmp_line) # Jouer la ligne tmp_line=[[cellule * 5],direction]
game.rep_lines.append(
tmp_line) # Ajouter la ligne à la liste des lignes jouées
game.calculer_lignes_jouables() # Recalcule des lignes jouables
best = 0
if (len(game.rep_lines) > best_score):
best_score = len(game.rep_lines) # Modifier le meilleur score
# Sauvegarder la liste des lignes du meilleur score
rep_lines_bestScore = game.rep_lines
best = 1
if (len(game.rep_playable_lines) == 0):
game.crash = True
else:
state_new = agent.get_state(game)
reward = agent.set_reward(game.crash, best)
#train short memory base on the new action and state
agent.train_short_memory(state_old, final_move, reward,
state_new, game.crash)
# store the new data into a long term memory
agent.remember(state_old, final_move, reward, state_new,
game.crash)
print("Score: ", len(game.rep_lines), " - Best score: ", best_score)
score_plot.append(len(game.rep_lines))
counter_plot.append(cpt_iteration)
cpt_iteration += 1
agent.replay_new(agent.memory)
score_ihm(max_iteration, best_score)
agent.model.save_weights('weights.hdf5')
plot_seaborn(counter_plot, score_plot)
afficher_lignes_ihm(rep_lines_bestScore)