noise = OUNoise(action_dim)
done = False
early_stop = False
train_epoch = 0
rewards = 0
while not early_stop:
# Sample trajectories
for step in range(NB_STEP):
action = model.actor.get_action(
torch.FloatTensor(state)
) * ACTION_BOUND # récupère l'action de l'agent et multiplie par la range de l'action
action = noise.get_action(
action.detach().numpy(),
step) # ajout de bruit pour l'exploration
next_state, reward, done, _ = env.step(
action) # avance dans l'environment
# On record les données de ce time_step
fifo.push(state, action, reward, next_state, done)
rewards += reward
state = next_state
if len(fifo) > BATCH_SIZE:
states_batch, actions_batch, rewards_batch, next_states_batch, done_batch = fifo.sample(
BATCH_SIZE)
states_batch = torch.FloatTensor(states_batch)
max_frames = 12000 * NUM_PROCESSES
max_steps = 500
frame_idx = 0
episode_rewards = []
batch_size = 128
if __name__ == "__main__":
# 초기 상태로 시작
while frame_idx < max_frames:
state = envs.reset()
ou_noise.reset()
episode_reward = 0
for step in range(max_steps):
action = policy_net.get_action(state)
action = ou_noise.get_action(action, step)
next_state, reward, done, _ = envs.step(action)
replay_buffer.push(state, action, reward, next_state, done)
if len(replay_buffer) > batch_size:
ddpg.update(batch_size, replay_buffer)
state = next_state
episode_reward += reward
frame_idx += NUM_PROCESSES
if frame_idx % (NUM_PROCESSES * 100) == 0:
# plot(frame_idx, rewards)
# rewards_tmp = np.array(rewards)
if episode_rewards:
print("finished frames {}, {:.1f}".format(