def main():
parameters = yaml.load(open(args.parameters_file, 'r'),
Loader=yaml.FullLoader)
model = Linear_QNet(11, 256, 3)
if args.use_trained == True:
model.load_state_dict(torch.load(parameters["model_path"]))
plot_scores = []
plot_mean_scores = []
total_score = 0
record = 0
agent = Agent(args, model)
game = SnakeGameAI()
while True:
# get old state
state_old = agent.get_state(game)
# get move
final_move = agent.get_action(state_old)
# perform move and get new state
reward, done, score = game.play_step(final_move)
state_new = agent.get_state(game)
# train short memory
agent.train_short_memory(state_old, final_move, reward, state_new,
done)
# remember
agent.remember(state_old, final_move, reward, state_new, done)
if done:
# train long memory, plot result
game.reset()
agent.n_games += 1
agent.train_long_memory()
if score > record:
record = score
if args.save_model == True:
agent.model.save()
print('Game', agent.n_games, 'Score', score, 'Record:', record)
plot_scores.append(score)
total_score += score
mean_score = total_score / agent.n_games
plot_mean_scores.append(mean_score)
plot(plot_scores, plot_mean_scores)
class Agent:
def __init__(self):
self.n_games = 0
self.epsilon = 0 # randomness
self.gamma = 0.9 # discount rate
self.memory = deque(maxlen = max_memory)
self.model = Linear_QNet(11, 256, 3)
PATH = './model/model.pth'
if os.path.exists(PATH):
self.model.load_state_dict(torch.load(PATH))
# self.model.eval()
print('Pretrained = True')
self.trainer = QTrainer(self.model, lr = lr, gamma = self.gamma)
def get_state(self, game):
head = game.snake[0]
point_l = Point(head.x - 20, head.y)
point_r = Point(head.x + 20, head.y)
point_u = Point(head.x, head.y - 20)
point_d = Point(head.x, head.y + 20)
dir_l = game.direction == Direction.LEFT
dir_r = game.direction == Direction.RIGHT
dir_u = game.direction == Direction.UP
dir_d = game.direction == Direction.DOWN
state = [
# Danger straight
(dir_r and game.is_collision(point_r)) or
(dir_l and game.is_collision(point_l)) or
(dir_u and game.is_collision(point_u)) or
(dir_d and game.is_collision(point_d)),
# Danger right
(dir_u and game.is_collision(point_r)) or
(dir_d and game.is_collision(point_l)) or
(dir_l and game.is_collision(point_u)) or
(dir_r and game.is_collision(point_d)),
# Danger left
(dir_d and game.is_collision(point_r)) or
(dir_u and game.is_collision(point_l)) or
(dir_r and game.is_collision(point_u)) or
(dir_l and game.is_collision(point_d)),
# Move direction
dir_l,
dir_r,
dir_u,
dir_d,
# Food location
game.food.x < game.head.x, # food left
game.food.x > game.head.x, # food right
game.food.y < game.head.y, # food up
game.food.y > game.head.y # food down
]
return np.array(state, dtype=int)
def remember(self, state, action, reward, next_state, done):
self.memory.append((state, action, reward, next_state, done))
def train_long_memory(self):
if len(self.memory) > batch_size:
mini_sample = random.sample(self.memory, batch_size) # list of tuples of size = 1000
else:
mini_sample = self.memory
states, actions, rewards, next_states, dones = zip(*mini_sample)
self.trainer.train_step(states, actions, rewards, next_states, dones)
def train_short_memory(self, state, action, reward, next_state, done):
self.trainer.train_step(state, action, reward, next_state, done)
def get_action(self, state):
# random moves: tradeoff exploration / exploitation
self.epsilon = 80 - self.n_games
final_move = [0, 0, 0]
if random.randint(0, 200) < self.epsilon:
move = random.randint(0, 2)
final_move[move] = 1
else:
state0 = torch.tensor(state, dtype = torch.float)
prediction = self.model(state0)
move = torch.argmax(prediction).item()
final_move[move] = 1
return final_move
class Agent:
def __init__(self, use_checkpoint=False):
self.no_of_games = 0
self.epsilon = 0 # randomness
self.gamma = 0.9 # discount rate
self.memory = deque(maxlen=MAX_MEMORY)
self.model = Linear_QNet(11, 256, 3)
self.trainer = QTrainer(self.model, lr=LR, gamma=self.gamma)
if use_checkpoint:
checkpoint = torch.load("./model/model.pth")
self.model.load_state_dict(checkpoint)
self.model.eval()
def get_state(self, game):
head = game.snake[0]
point_l = Point(head.x - BLOCK_SIZE, head.y)
point_r = Point(head.x + BLOCK_SIZE, head.y)
point_u = Point(head.x, head.y - BLOCK_SIZE)
point_d = Point(head.x, head.y + BLOCK_SIZE)
dir_l = game.direction == Direction.LEFT
dir_r = game.direction == Direction.RIGHT
dir_u = game.direction == Direction.UP
dir_d = game.direction == Direction.DOWN
state = [
# Danger straight
(dir_r and game.is_collision(point_r))
or (dir_l and game.is_collision(point_l))
or (dir_u and game.is_collision(point_u))
or (dir_d and game.is_collision(point_d)),
# Danger right
(dir_u and game.is_collision(point_r))
or (dir_d and game.is_collision(point_l))
or (dir_l and game.is_collision(point_u))
or (dir_r and game.is_collision(point_d)),
# Danger left
(dir_d and game.is_collision(point_r))
or (dir_u and game.is_collision(point_l))
or (dir_r and game.is_collision(point_u))
or (dir_l and game.is_collision(point_d)),
# Move direction
dir_l,
dir_r,
dir_u,
dir_d,
# Food location
game.food.x < game.head.x, # Food left
game.food.x > game.head.x, # Food right
game.food.y < game.head.y, # Food up
game.food.y > game.head.y, # Food down
]
return np.array(state, dtype=int)
def remember(self, state, action, reward, next_state, game_over):
self.memory.append((state, action, reward, next_state, game_over))
def train_long_memory(self):
if len(self.memory) > BATCH_SIZE:
mini_sample = random.sample(self.memory, BATCH_SIZE)
else:
mini_sample = self.memory
states, actions, rewards, next_states, game_overs = zip(*mini_sample)
self.trainer.train_step(states, actions, rewards, next_states,
game_overs)
def train_short_memory(self, state, action, reward, next_state, game_over):
self.trainer.train_step(state, action, reward, next_state, game_over)
def get_action(self, state):
self.epsilon = 80 - self.no_of_games
action = [0, 0, 0]
if random.randint(0, 200) < self.epsilon:
move = random.randint(0, 2)
action[move] = 1
else:
state0 = torch.tensor(state, dtype=torch.float)
prediction = self.model(state0)
move = torch.argmax(prediction).item()
action[move] = 1
return action
class Agent:
def __init__(self):
self.n_games = 0
self.epsilon = 0 #randomness
self.gamma = 0.9 #discount rate
self.memory = deque(maxlen=MAX_MEMORY) #popleft()
self.model = Linear_QNet(11, 256,
3) #input_lauer=11,hidden:256 ,output:3
self.model.load_state_dict(torch.load('./optimized_model/model.pth'))
self.trainer = QTrainer(self.model, lr=LR, gamma=self.gamma)
def get_state(self, game):
head = game.snake[0]
BLOCK_SIZE = 20
#Points to check danger
point_l = Point(head.x - BLOCK_SIZE, head.y)
point_r = Point(head.x + BLOCK_SIZE, head.y)
point_u = Point(head.x, head.y - BLOCK_SIZE)
point_d = Point(head.x, head.y + BLOCK_SIZE)
dir_l = game.direction == Direction.LEFT
dir_r = game.direction == Direction.RIGHT
dir_u = game.direction == Direction.UP
dir_d = game.direction == Direction.DOWN
state = [
#For straight
(dir_r and game.is_collision(point_r))
or (dir_l and game.is_collision(point_l))
or (dir_u and game.is_collision(point_u))
or (dir_d and game.is_collision(point_d)),
#Danger Right
(dir_u and game.is_collision(point_r))
or (dir_d and game.is_collision(point_l))
or (dir_l and game.is_collision(point_u))
or (dir_r and game.is_collision(point_d)),
#Danger left
(dir_d and game.is_collision(point_r))
or (dir_u and game.is_collision(point_l))
or (dir_r and game.is_collision(point_u))
or (dir_l and game.is_collision(point_d)),
#Move direction
dir_l,
dir_r,
dir_u,
dir_d,
#Food location
game.food.x < game.head.x, # food left
game.food.x > game.head.x, # food right
game.food.y < game.head.y, # food up
game.food.y > game.head.y # food down
]
return np.array(state, dtype=int)
def remember(self, state, action, reward, next_state, done):
self.memory.append((state, action, reward, next_state,
done)) # popleft if MAX_MEMORY IS REACHED
def train_long_memory(self):
if len(self.memory) > BATCH_SIZE:
mini_sample = random.sample(self.memory,
BATCH_SIZE) #list of tuples
else:
mini_sample = self.memory
states, actions, rewards, next_states, dones = zip(*mini_sample)
self.trainer.train_step(states, actions, rewards, next_states, dones)
#for state, action,reward, next_state, done in mini_sample:
# self.trainer.train_step(state, action,reward, next_state, done)
def train_short_memory(self, state, action, reward, next_state, done):
self.trainer.train_step(state, action, reward, next_state, done)
def get_action(self, state):
# random moves : tradeoff exploration / exploitation
self.epsilon = 80 - self.n_games
final_move = [0, 0, 0]
if random.randint(
0, 200
) < self.epsilon and False: #This was original ,we made small changes to it
#if random.randint(0,200) < 20 and self.n_games<90:
move = random.randint(0, 2)
final_move[move] = 1
else:
state0 = torch.tensor(state, dtype=torch.float)
prediction = self.model.forward(state0)
move = torch.argmax(prediction).item()
final_move[move] = 1
return final_move
class Agent:
def __init__(self):
self.numberOfGames = 0
self.epsilon = 0 # controlls randomness
self.gamma = 0.9 # discount rate, <1
# will popleft if there is too much in memory
self.memory = deque(maxlen=maxMemory)
self.model = Linear_QNet(11, 256, 3)
if os.path.isfile('./model/model.pth'):
model_folder_path = './model/model.pth'
self.model.load_state_dict(torch.load(model_folder_path))
self.trainer = QTrainer(self.model, lr=learningRate, gamma=self.gamma)
def getState(self, game):
head = game.snake[0]
# Clok-wise directions and angles
cw_dirs = [
Direction.RIGHT == game.direction,
Direction.DOWN == game.direction, Direction.LEFT == game.direction,
Direction.UP == game.direction
]
cw_angs = np.array([0, np.pi / 2, np.pi, -np.pi / 2])
# Position - in front: 0, on right: 1, on left: -1; BLOCK_SIZE = 20
def getPoint(pos):
return Point(
head.x + 20 * np.cos(cw_angs[(cw_dirs.index(True) + pos) % 4]),
head.y + 20 * np.sin(cw_angs[(cw_dirs.index(True) + pos) % 4]))
state = [
# Danger
game.is_collision(getPoint(0)),
game.is_collision(getPoint(1)),
game.is_collision(getPoint(-1)),
# Move direction
cw_dirs[2],
cw_dirs[0],
cw_dirs[3],
cw_dirs[1],
# Food location
game.food.x < head.x,
game.food.x > head.x,
game.food.y < head.y,
game.food.y > head.y
]
return np.array(state, dtype=int)
def remember(self, state, action, reward, next_state, game_over):
self.memory.append((state, action, reward, next_state, game_over))
def trainLongMemory(self):
if len(self.memory) > batchSize:
# list of tuples from the memory
miniSample = random.sample(self.memory, batchSize)
else:
miniSample = self.memory
states, actions, rewards, next_states, game_over = zip(*miniSample)
self.trainer.trainStep(states, actions, rewards, next_states,
game_over)
def trainShortMemory(self, state, action, reward, next_state, game_over):
self.trainer.trainStep(state, action, reward, next_state, game_over)
def getAction(self, state):
# exploitation / exploration
self.epsilon = 80 - self.numberOfGames
final_move = [0, 0, 0]
if random.randint(-2, 200) < self.epsilon:
move = random.randint(0, 2)
final_move[move] = 1
else:
state0 = torch.tensor(state, dtype=torch.float)
prediction = self.model(state0)
move = torch.argmax(prediction).item()
final_move[move] = 1
return (final_move)