from ai import DQN
Config.set('input', 'mouse', 'mouse, multitouch_on_demand') # Adding this line if we don't want the right click to put a red point
# Introducing last_signal1, last_signal2 and last_signal3 to let the brain know if it's getting further or closer to the obstacle.
last_signal1 = 0
last_signal2 = 0
last_signal3 = 0
last_x = 0
last_y = 0
GOAL = 'airport'
BOUNDARY = 20
# Getting our AI, which we call "dqn", and that contains our neural network that represents our Q-function
dqn = DQN(8, 3, 0.9) # 5 signals, 3 actions, gamma = 0.9
action2rotation = [0, 10, -10] # action = 0: no rotation, action = 1, rotate 10 degrees, action = 2, rotate -10 degrees
last_reward = 0 # initializing the last reward
scores = [] # initializing the mean score curve (sliding window of the rewards) w.r.t time
# Initializing the map
first_update = True # using this trick to initialize the map only once
def init():
global sand # sand is an array that has as many cells as our graphic interface has pixels. Each cell has a one if there is sand, 0 otherwise.
global goal_x # x-coordinate of the goal (where the car has to go, that is the up-left corner or the bot-right corner)
global goal_y # y-coordinate of the goal (where the car has to go, that is the up-left corner or the bot-right corner)
global first_update
sand = np.zeros((RIGHT, TOP)) # initializing the sand array with only zeros
goal_x = 30 # the goal to reach is at the upper left of the map (the x-coordinate is 20 and not 0 because the car gets bad reward if it touches the wall)
goal_y = TOP - 30 # the goal to reach is at the upper left of the map (y-coordinate)
first_update = False # trick to initialize the map only once
from kivy.clock import Clock
# Importing the DQN object from our AI in ai.py
from ai import DQN
# Adding this line if we don't want the right click to put a red point
Config.set('input', 'mouse', 'mouse,multitouch_on_demand')
# Introducing last_x and last_y, used to keep the last point in memory when we draw the sand on the map
last_x = 0
last_y = 0
n_points = 0
length = 0
# Getting our AI, which we call "brain", and that contains our neural network that represents our Q-function
brain = DQN(5, 3, 0.9)
action2rotation = [0, 20, -20]
last_reward = 0
scores = []
# Initializing the map
first_update = True
def init():
global sand
global goal_x
global goal_y
global first_update
sand = np.zeros((longueur, largeur))
goal_x = 20
# Importing the Dqn object from our AI in ai.py
# Basically this is the brain of the AI in the car --> from the Ai.py file
# Dqn stands for deep q network
from ai import DQN
# Adding this line if we don't want the right click to put a red point
Config.set('input', 'mouse', 'mouse,multitouch_on_demand')
# Introducing last_x and last_y, used to keep the last point in memory when we draw the sand on the map
last_x = 0
last_y = 0
n_points = 0 # the total number of points in the last drawing
length = 0 # the length of the last drawing
# Getting our AI, which we call "brain", and that contains our neural network that represents our Q-function
brain = DQN(5, 3, 0.95) # 5 sensors, 3 actions, gama = 0.9
action2rotation = [
0, 20, -20
] # action = 0 => no rotation, action = 1 => rotate 20 degres, action = 2 => rotate -20 degres
last_reward = 0 # initializing the last reward
scores = [
] # initializing the mean score curve (sliding window of the rewards) with respect to time
# Initializing the map
first_update = True # using this trick to initialize the map only once
def init():
global sand # sand is an array that has as many cells as our graphic interface has pixels. Each cell has a one if there is sand, 0 otherwise.
global goal_x # x-coordinate of the goal (where the car has to go, that is the airport or the downtown)
global goal_y # y-coordinate of the goal (where the car has to go, that is the airport or the downtown)
input_size = GAME.board_state.shape
action_space_size = 32 * 32
conv_layer_sizes = [(128, 2, 1), (128, 2, 1), (128, 2, 1)]
hidden_layer_sizes = [256]
gamma = 0.99
batch_sz = 32
num_episodes = 10000
total_t = 0
experience_replay_buffer = ReplayMemory(input_size)
episode_rewards = np.zeros(num_episodes)
epsilon = 0.001
model1 = DQN(input_size, action_space_size, conv_layer_sizes,
hidden_layer_sizes, 'model1')
model2 = DQN(input_size, action_space_size, conv_layer_sizes,
hidden_layer_sizes, 'model2')
with tf.Session() as session:
model1.set_session(session)
model2.set_session(session)
session.run(tf.global_variables_initializer())
GAME.reset()
wins = 0
for i in range(MIN_EXPERIENCES):
GAME.available_moves()
if GAME.win != 0:
GAME.reset()
move = random_play(GAME)
action = encoding_move(move)