def train(game):
agent = DQN(game)
for i in tqdm(range(TRAIN_GAMES)):
game.new_episode()
previous_variables = None
previous_img = None
done = False
local_history = []
total_reward = 0
while not done:
state = game.get_state()
img = state.screen_buffer
variables = state.game_variables
if previous_variables is None:
previous_variables = variables
if previous_img is None:
previous_img = img
action = agent.act(img)
reward = game.make_action(action)
done = game.is_episode_finished()
reward = (reward + calculate_additional_reward(previous_variables, variables)) / 100
total_reward += reward
local_history.append([previous_img, img, reward, action, done])
previous_variables = variables
previous_img = img
if total_reward >= 0:
for previous_state, state, reward, action, done in local_history:
agent.remember(previous_state, state, reward, action, done)
agent.train()
def run(ep,train=False):
pygame.init()
loss=[]
agent = DQN(3, 5)
env=pongGame()
weights_filepath = 'PongGame.h5'
if train==False:
agent.model.load_weights(weights_filepath)
print("weights loaded")
for e in range(ep):
for event in pygame.event.get():
if event.type == pygame.QUIT:
pygame.quit()
quit()
state = env.reset()
state = np.reshape(state, (1, 5))
score = 0
max_steps = 1000
for i in range(max_steps):
action = agent.act(state)
reward, next_state, done = env.step(action)
score += reward
next_state = np.reshape(next_state, (1, 5))
agent.remember(state, action, reward, next_state, done)
state = next_state
if train==True:
agent.replay()
if done:
print("episode: {}/{}, score: {}".format(e, ep, score))
break
loss.append(score)
if train:
agent.model.save_weights("PongGame.h5")
return loss
def play(game):
agent = DQN(game, use_saved=True)
for i in tqdm(range(PLAY_GAMES)):
game.new_episode()
done = False
while not done:
state = game.get_state()
img = state.screen_buffer
action = agent.act(img)
print(action)
game.make_action(action)
done = game.is_episode_finished()
N_SAVE = 500
env = gym.make('FlappyBird-v0')
agent = DQN(env)
scores = deque(maxlen=100)
for i in range(N_EP):
score = 0
ob = env.reset()
# Stack observations
pre_ob = preprocess(ob)
pre_ob = pre_ob.reshape(1, 100, 100)
ob_stack = np.stack((pre_ob, ) * 4, -1)
pre_ob = ob_stack
while True:
action = agent.act(pre_ob, step=i)
ob, reward, done, _ = env.step(action)
if reward <= -1:
reward = -1
next_pre_ob = preprocess(ob)
# Stack observations
next_pre_ob = next_pre_ob.reshape(1, 100, 100)
ob_stack = np.insert(ob_stack, -1, next_pre_ob, axis=3)
ob_stack = np.delete(ob_stack, 0, axis=3)
next_pre_ob = ob_stack
agent.remember(pre_ob, action, reward, next_pre_ob, done)
agent.replay()