params.log_dir = "../logs/logs/{}".format(params.env_name)
params.model_dir = "../logs/models/{}".format(params.env_name)
env = gym.make(params.env_name)
# set seed
env.seed(params.seed)
tf.compat.v1.random.set_random_seed(params.seed)
agent = DDPG(Actor, Critic, env.action_space.shape[0], params)
replay_buffer = ReplayBuffer(params.memory_size)
reward_buffer = deque(maxlen=params.reward_buffer_ep)
summary_writer = tf.contrib.summary.create_file_writer(params.log_dir)
init_state = env.reset() # reset
agent.predict(init_state) # burn the format of the input matrix to get the weight matrices!!
gp_model, update = create_bayes_net()
optimiser = tf.compat.v1.train.AdamOptimizer()
num_sample = 100 # number of sampling
get_ready(agent.params)
global_timestep = tf.compat.v1.train.get_or_create_global_step()
time_buffer = deque(maxlen=agent.params.reward_buffer_ep)
log = logger(agent.params)
with summary_writer.as_default():
# for summary purpose, we put all codes in this context
with tf.contrib.summary.always_record_summaries():
policies, scores = list(), list()
for i in itertools.count():
with tf.contrib.summary.always_record_summaries():
for i in itertools.count():
state = env.reset()
total_reward = 0
self_rewards = 0
start = time.time()
agent.random_process.reset_states()
done = False
episode_len = 0
while not done:
traj.append(state)
if global_timestep.numpy() < agent.params.learning_start:
action = env.action_space.sample()
else:
action = agent.predict(state)
next_state, reward, done, info = env.step(
np.clip(action, -1.0, 1.0))
replay_buffer.add(state, action, reward, next_state, done)
"""
=== Update the models
"""
if global_timestep.numpy() > agent.params.learning_start:
states, actions, rewards, next_states, dones = replay_buffer.sample(
agent.params.batch_size)
loss = agent.update(states, actions, rewards, next_states,
dones)
soft_target_model_update_eager(
agent.target_actor,
agent.actor,
tau=agent.params.soft_update_tau)