def main(args, reward_result):
with tf.Session() as sess:
# Initialize environment and seed
np.random.seed(int(args['random_seed']))
tf.set_random_seed(int(args['random_seed']))
env = allCars()
state_dim = 6
action_dim = 1
action_bound = 5
actor = ActorNetwork(sess, state_dim, action_dim, action_bound,
float(args['actor_lr']), float(args['tau']),
int(args['minibatch_size']))
critic = CriticNetwork(sess, state_dim, action_dim,
float(args['critic_lr']), float(args['tau']),
float(args['gamma']),
actor.get_num_trainable_vars())
actor_noise = OrnsteinUhlenbeckActionNoise(mu=np.zeros(action_dim))
[summary_ops, summary_vars,
paths] = train(sess, env, args, actor, critic, actor_noise,
reward_result)
return [summary_ops, summary_vars, paths]
def main(args, reward_result):
with tf.Session() as sess:
# Initialize environment and seed
np.random.seed(int(args['random_seed']))
tf.set_random_seed(int(args['random_seed']))
env = allCars()
state_dim = 6
action_dim = 1
action_bound = 5
learner = Learner(sess, env, state_dim, action_dim, action_bound,
float(args['actor_lr']), float(args['critic_lr']),
float(args['tau']), float(args['gamma']),
float(args['minibatch_size']))
#[summary_ops, summary_vars, paths, reward_result] = learner.train(replay_buffer, minibatch_size)
[summary_ops, summary_vars, paths,
reward_result] = learner.train_rollout(args, reward_result)
return [summary_ops, summary_vars, paths]
# Continuous environments
ENVIRONMENT = 'Car-Data'
# ENVIRONMENT = 'Pendulum-v0'
# ENVIRONMENT = 'CartPole-v1'
# ENVIRONMENT = 'MountainCarContinuous-v0'
# ENVIRONMENT = 'LunarLanderContinuous-v2'
# ENVIRONMENT = 'BipedalWalker-v2'
# ENVIRONMENT = 'BipedalWalkerHardcore-v2'
# ENVIRONMENT = 'CarRacing-v0'
TIMESTAMP = datetime.now().strftime("%Y%m%d-%H%M%S")
SUMMARY_DIR = os.path.join(OUTPUT_RESULTS_DIR, "PPO", ENVIRONMENT,
TIMESTAMP)
#env = gym.make(ENVIRONMENT)
env = allCars()
#env = wrappers.Monitor(env, os.path.join(SUMMARY_DIR, ENVIRONMENT), video_callable=None)
ppo = PPO(env, SUMMARY_DIR, gpu=True)
if MODEL_RESTORE_PATH is not None:
ppo.restore_model(MODEL_RESTORE_PATH)
t, terminal = 0, False
buffer_s, buffer_a, buffer_r, buffer_v, buffer_terminal = [], [], [], [], []
rolling_r = RunningStats()
# Get prior and set tuning parameters for adaptive regularization weight
prior = BasePrior()
lambda_store = np.zeros(BATCH + 1)
lambda_all = np.zeros(EP_MAX + 1)
lambda_max = 8