class Agent:
def __init__(self, load_path=''):
self.n_games = 0
self.epsilon = 0
self.gamma = 0.9
self.load_path = load_path
self.memory = deque(maxlen=MAX_MEMORY)
self.model = Net(11, 256, 3)
if load_path:
self.model.load_state_dict(torch.load(load_path))
self.trainer = Trainer(self.model, LR, self.gamma)
def get_state(self, game):
# 0 1 2 3
# U L R D
# [[1, -10], [0, -10], [0, 10], [1, 10]]
head = game.snake_pos
near_head = [
[head[0], head[1] - 10],
[head[0] - 10, head[1]],
[head[0] + 10, head[1]],
[head[0], head[1] + 10],
]
directions = [
game.direction == 0,
game.direction == 1,
game.direction == 2,
game.direction == 3,
]
state = [
(directions[0] and game.is_colision(near_head[0])) or
(directions[1] and game.is_colision(near_head[1])) or
(directions[2] and game.is_colision(near_head[2])) or
(directions[3] and game.is_colision(near_head[3])),
(directions[0] and game.is_colision(near_head[1])) or
(directions[1] and game.is_colision(near_head[3])) or
(directions[2] and game.is_colision(near_head[0])) or
(directions[3] and game.is_colision(near_head[2])),
(directions[0] and game.is_colision(near_head[2])) or
(directions[1] and game.is_colision(near_head[0])) or
(directions[2] and game.is_colision(near_head[3])) or
(directions[3] and game.is_colision(near_head[1])),
game.food_pos[0] < head[0],
game.food_pos[0] > head[0],
game.food_pos[1] < head[1],
game.food_pos[1] > head[1],
] + directions
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)
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):
if not self.load_path:
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):
self.num_games = 0 # number of games played
self.epsilon = 0 # randomness
self.gamma = 0.9 # discount rate
self.memory = deque(
maxlen=MAX_MEMORY) # pops from left if memory limit is exceeded
self.model = Linear(11, 256, 3)
self.trainer = Trainer(self.model, lr=LEARNING_RATE, 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)
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):
# initially the agent performs more of random moves i.e exploration
self.epsilon = 80 - self.num_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
