def generate_data(DQN, min_epsilon, epsilon, copy_step):
maze = game.MazeGame(9,7)
maze.create_path(20)
maze.draw_path()
maze.clean_path()
is_finished = False
reward = 0
turn = 0
decay = 0.999995
fails = 0
repetitions = 1
loss = 0
while repetitions % 2 != 0:
prev_observations = maze.return_state_2d()
possible_actions = maze.possible_actions()
action, action_key = DQN.get_action(prev_observations, epsilon, possible_actions)
maze.move(action)
is_finished, reward = maze.check(turn)
observations = maze.return_state_2d()
turn += 1
exp = {'prev_obs': prev_observations, 'a' : action_key-1, 'r': reward, 'obs': observations, 'done' : is_finished }
DQN.add_experience(exp)
if is_finished:
maze.reset() # TODO
repetitions +=1
is_finished = False
if turn % 20 == 0:
epsilon = max(epsilon*decay, min_epsilon)
if turn % 8 == 0:
loss += DQN.train()
if turn % 40 == 0:
fails = 1
return fails, turn, loss, epsilon
return fails, turn, loss, epsilon
def __init__(self, handle):
"""Set up the Maze activity."""
activity.Activity.__init__(self, handle)
self._busy_count = 0
self._unbusy_idle_sid = None
self.build_toolbar()
self.pservice = PresenceService()
self.owner = self.pservice.get_owner()
state = None
if 'state' in self.metadata:
state = json.loads(self.metadata['state'])
self.game = game.MazeGame(self, self.owner, state)
self.set_canvas(self.game)
self.game.show()
self.connect("key_press_event", self.game.key_press_cb)
self.text_channel = None
self.my_key = profile.get_pubkey()
self._alert = None
if self.shared_activity:
# we are joining the activity
self._add_alert(_('Joining a maze'), _('Connecting...'))
self.connect('joined', self._joined_cb)
if self.get_shared():
# we have already joined
self._joined_cb()
else:
# we are creating the activity
self.connect('shared', self._shared_cb)
def __init__(self, maze):
pygame.init()
pygame.font.init()
self.DQN = MazeDQN.MazeDQN(63, 4, 9, 7)
self.DQN.load_weights("mazenet")
self.side_length = 100
self.maze = maze
self.done = False
self.height = maze.rows
self.width = maze.col
self.imagerect = (0, 0)
self.black = (0, 0, 0)
self.white = (255, 255, 255)
self.red = (255, 0, 0)
self.green = (0, 255, 0)
self.open = (125, 125, 125)
self.blue = (0, 0, 125)
self.screen = pygame.display.set_mode([
self.side_length * 2 + self.width * 100,
self.side_length + 100 * self.height
])
self.text = pygame.font.SysFont('Arial Black', 30)
while True:
maze = self.maze
self.mouse_pos = pygame.mouse.get_pos()
self.render(self.white, self.black, self.red, self.green,
self.white)
for event in pygame.event.get():
if event.type == pygame.QUIT:
pygame.quit()
exit()
if event.type == pygame.MOUSEBUTTONUP:
prev_observations = maze.return_state_2d()
possible_actions = maze.possible_actions()
action, action_key = self.DQN.get_action(
prev_observations, 0, possible_actions)
position = maze.game_move(action)
print(maze.board)
print(action)
is_finished, reward = maze.check(0)
if is_finished:
self.done = True
self.render(self.white, self.black, self.red,
self.green, self.white)
sleep(4)
maze = game.MazeGame(9, 7)
maze.create_path(20)
self.maze = maze
self.done = False
def play_game(DQN):
maze = game.MazeGame(9,7)
maze.create_path(20)
maze.draw_path()
maze.render()
is_finished = False
reward = 0
turn = 0
while not is_finished:
possible_actions = maze.possible_actions()
action, action_key = DQN.get_action(maze.board, 0.9, possible_actions)
maze.move(action)
is_finished, reward = maze.check(turn)
maze.clean_render()
turn += 1
def dojo(DQN, iterations, min_epsilon, epsilon, copy_step):
total_loss = 0
total_fails = 0
total_turns = 0
games = 1
decay = 0.99995
test_game = game.MazeGame(9,7)
test_state = test_game.return_state_2d()
test_predict = DQN.predict(np.atleast_2d(test_state)).detach().numpy()
print(test_predict)
for i in range(iterations):
fails, turns, loss, epsilon = generate_data(DQN, min_epsilon, epsilon, copy_step)
loss += DQN.train()
total_fails += fails
total_loss += loss
total_turns += turns
games +=1
if i % 100 == 0 and i != 0:
print("total loss:", total_loss)
print("average turns:", total_turns/ games)
print("average fails", total_fails / games)
print("epsilon", epsilon)
games = 0
total_fails = 0
total_loss = 0
total_turns = 0
print("games", i)
if i % 8 == 0:
DQN.copy_weights()
test_predict = DQN.predict(np.atleast_2d(test_state)).detach().numpy()
print(test_predict)
if i % 10 == 0:
epsilon = max(epsilon*decay, min_epsilon)
if i % 1000 == 0 and i != 0:
plot = plot_grad_flow(DQN.model.named_parameters())
path = "plot" + str(i)+ ".png"
plot.savefig(path)
DQN.save_weights("mazenet")
else:
textsurface = self.text.render("FERDIG!", True, (0, 0, 0))
self.screen.blit(textsurface, (200, 0))
#vertical lines
for line in range(self.width + 1):
pygame.draw.line(
self.screen, self.black, [line * 300 + self.side_length, 50],
[line * 300 + self.side_length, self.height * 100 + 50], 3)
for line in range(self.height + 1):
pygame.draw.line(
self.screen, self.black, [self.side_length, line * 300 + 50],
[self.side_length + self.width * 100, line * 300 + 50], 3)
for col in range(self.width):
for row in range(self.height):
space = (col, row)
if self.task.board[space] == 1:
pygame.draw.rect(self.screen, self.blue,
(self.side_length + 1 + col * 100,
51 + row * 100, 98, 98))
if self.task.board[space] == -1:
pygame.draw.rect(self.screen, self.red,
(self.side_length + 1 + col * 100,
51 + row * 100, 98, 98))
pygame.event.pump()
pygame.display.flip()
task = game.MazeGame(9, 7)
task.create_path(20)
render(task)