def getTrainingData(self):
print('Getting Training Data . . .')
data = []
number = int(self.train_games / 20)
for x in range(self.train_games):
game = Game(x=self.x, y=self.y)
c_data = []
self.game = game
snake = game.start()
current_state = self.getState(snake)
for _ in range(self.max_steps):
action = self.getAction()
length = snake.length
done, snake, closer = game.step(action)
if done: break
elif not closer: continue
else:
correct_output = [0, 0, 0]
correct_output[action + 1] = 1
num = 1
if snake.length > length: num = 3
for _ in range(num):
c_data.append([current_state, correct_output])
current_state = self.getState(snake)
if snake.length > 2:
for el in c_data:
data.append(el)
if x % number == 0: print(f'{int(x/self.train_games*100)}%')
return data
def eval_genome(genome, config):
net = neat.nn.FeedForwardNetwork.create(genome, config)
fitnesses = []
for runs in range(runs_per_net):
game = Game(20, 20)
# Run the given simulation for up to num_steps time steps.
fitness = 0.0
while True:
inputs = game.get_normalized_state()
action = net.activate(inputs)
# Apply action to the simulated snake
valid = game.step(np.argmax(action))
# Stop if the network fails to keep the snake within the boundaries or hits itself.
# The per-run fitness is the number of pills eaten
if not valid:
break
fitness = game.fitness
fitnesses.append(fitness)
# The genome's fitness is its worst performance across all runs.
return min(fitnesses)
def eval_genome(genome, config):
net = neat.nn.FeedForwardNetwork.create(genome, config)
fitnesses = []
for runs in range(runs_per_net):
#pygame.init()
#screen = pygame.display.set_mode((20 * 16,20 * 16))
#screen.fill(pygame.Color('black'))
#pygame.display.set_caption('Snake')
#pygame.display.flip()
sim = Game(20, 20)
# Run the given simulation for up to num_steps time steps.
fitness = 0.0
while True:
inputs = sim.get_normalized_state()
action = net.activate(inputs)
# Apply action to the simulated snake
valid = sim.step(np.argmax(action))
# Stop if the network fails to keep the snake within the boundaries or hits itself.
# The per-run fitness is the number of pills eaten
if not valid:
break
fitness = sim.score
fitnesses.append(fitness)
# The genome's fitness is its worst performance across all runs.
return min(fitnesses)
class SnakeWrapper:
"""
return the croped square_size-by-square_size after rotation and changing to one-hot and doing block-notation.
"""
# num_classes is the number of different element types that can be found on the board.
# yes I know, actually we have 9 types, but 10 is nicer. (4 snakes + 1 obstacle + 3 fruits + 1 empty = 9)
num_classes = 10
# the action space. 0-left, 1-forward, 2-right.
action_space = gym.spaces.Discrete(3)
# the observation space. 9x9 one hot vectors, total 9x9x10.
# your snake always look up (the observation is a rotated crop of the board).
observation_space = gym.spaces.Box(
low=0,
high=num_classes,
shape=(9, 9, 10),
dtype=np.int
)
def __init__(self):
self.game = Game()
self.square_size = 9 # the observation size
self.timestep = 0
def step(self, action):
# get action as integer, move the game one step forward
# return tuple: state, reward, done, info. done is always False - Snake game never ends.
action = int_to_action[action]
reward = self.game.step(action)
head_pos = self.game.players[1].chain[-1]
direction = self.game.players[1].direction
board = self.game.board
state = preprocess_snake_state(board, head_pos, direction, self.square_size, SnakeWrapper.num_classes)
self.timestep += 1
return state, reward
def seed(self, seed=None):
return self.game.seed(seed)
# reset the game and return the board observation
def reset(self):
self.game.reset()
self.timestep = 0
first_state, _ = self.step(0)
return first_state
# print the board to the console
def render(self, mode='human'):
self.game.render(self.timestep)
def showGame(self, model):
game = Game(x=self.x, y=self.y, gui=True)
self.game = game
while True:
snake = game.start()
steps = self.max_steps
current_state = self.getState(snake)
while True:
m = model.predict(np.array([current_state]))
action = list(m[0]).index(max(list(m[0]))) - 1
length = snake.length
done, snake, c = game.step(action)
if done: break
elif snake.length > length: steps = self.max_steps
else: current_state = self.getState(snake)
time.sleep(.05)
steps -= 1
if steps == 0:
break
def test(self, model):
print('Testing . . .')
num = int(self.test_games / 20)
lengths = []
game = Game(x=self.x, y=self.y)
self.game = game
for x in range(self.test_games):
snake = game.start()
steps = self.max_steps
current_state = self.getState(snake)
while True:
m = model.predict(np.array([current_state]))
action = list(m[0]).index(max(list(m[0]))) - 1
length = snake.length
done, snake, _ = game.step(action)
if done: break
elif snake.length > length: steps = self.max_steps
else: current_state = self.getState(snake)
steps -= 1
if steps == 0:
break
lengths.append(snake.length)
if x % num == 0: print(f'{int((x/self.test_games)*100)}%')
print(f'Average: {sum(lengths)/len(lengths)}')
game = Game(food_ammount=1, render=True)
valid = True
observation = Game().reset()
score = 0
q_table = np.load(f"{FILE}.npy", allow_pickle=True)
os.makedirs(f"{FILE}", exist_ok=True)
step = 0
while valid:
game.draw()
surface = pygame.display.get_surface()
pygame.image.save(surface, f"{FILE}/image_{step}.png")
old_observation = observation
current_q_values = get_discrete_vals(q_table, old_observation)
action = np.argmax(current_q_values)
old_q = current_q_values[action]
valid, reward, observation = game.step(action=action)
step += 1
# time.sleep(0.03)
game.draw()
surface = pygame.display.get_surface()
pygame.image.save(surface, f"{FILE}/image_{step}.png")
