def main():
torch.cuda.set_device(0)
seed_num = 1
torch.cuda.manual_seed(seed_num)
# data_dir = '/home/bike/data/mnist/'
out_dir = '/home/becky/Git/reinforcement_learning_pytorch/log/REINFORCEMENT_{}/'.format(
datetime.now())
if not os.path.exists(out_dir):
os.makedirs(out_dir)
shutil.copyfile(sys.argv[0], out_dir + '/REINFORCE_cart_pole.py')
sys.stdout = logger.Logger(out_dir)
env_name = 'CartPole-v0'
killer = GracefulKiller()
env, obs_dim, act_dim = init_gym(env_name)
num_episodes = 300
rewards = np.zeros(num_episodes)
QValue = REINFORCEAgent(obs_dim,
act_dim,
learning_rate=0.0001,
reward_decay=0.99,
e_greedy=0.9)
for i_episode in range(num_episodes):
rewards[i_episode] = run_policy(env, QValue, episodes=100)
print("In episode {}, the reward is {}".format(
str(i_episode), str(rewards[i_episode])))
if killer.kill_now:
now = "REINFORCE_v1"
QValue.save_model(str(now))
break
print('game over!')
util.before_exit(model=QValue.model, reward=rewards)
env.close()
env.render(close=True)
g = 0
for t in reversed(range(len(rewards))):
g = rewards[t] + gamma * g
returns[t] = g
trajectory['returns'] = returns
seed = 0
env = REINFORCEEnv()
np.random.seed(seed)
tf.set_random_seed(seed)
env.seed(seed=seed)
obs_dim = env.observation_space.shape[0]
n_act = 7 #config: act_dim #env.action_space.n
agent = REINFORCEAgent(obs_dim, n_act, epochs=5, hdim=32, lr=3e-4, seed=seed)
avg_return_list = deque(maxlen=1000)
avg_loss_list = deque(maxlen=1000)
episode_size = 1
batch_size = 16
nupdates = 10000
for update in range(nupdates + 1):
#print ('update: ', update)
trajectories = run_policy(env, agent, episodes=episode_size)
compute_returns(trajectories)
observes, actions, returns = build_train_set(trajectories)
pol_loss = agent.update(observes, actions, returns, batch_size=batch_size)