state = next_state
episode_reward += reward
if done:
args_k = k_by_reward(episode_reward)
state = env.reset()
all_rewards.append(episode_reward)
episode_reward = 0
if len(replay_buffer) > batch_size:
loss = improved_td_loss(batch_size, replay_buffer,
current_model, target_model, gamma,
args_k, optimizer)
losses.append(loss.data[0])
if frame_idx % 1000 == 0:
update_target(current_model, target_model)
if frame_idx % 1000 == 0:
print("This is %d frame of 1000000, k_end %f, reward is %f" %
(frame_idx, k_end * 2., all_rewards[-1]))
losses_all.append(losses)
rewards_all.append(all_rewards)
str = "data/analysis/CartPolerl1_kend%f" % (k_end * 2.)
mean_losses, var_losses = tl.StatShrink2D(losses_all)
mean_rewards, var_rewards = tl.StatShrink2D(rewards_all)
tl.save2D4list(mean_losses, var_losses, mean_rewards, var_rewards, str)
next_state, reward, done, _ = env.step(action)
replay_buffer.push(state, action, reward, next_state, done)
state = next_state
episode_reward += reward
if done:
state = env.reset()
all_rewards.append(episode_reward)
episode_reward = 0
if len(replay_buffer) > batch_size:
loss = compute_td_loss1(batch_size, replay_buffer, current_model,
target_model, gamma, optimizer)
losses.append(loss.data[0])
if frame_idx % 1000 == 0:
update_target(current_model, target_model)
if frame_idx % 1000 == 0:
print("This is %d frame of 1000000, range %d, reward is %f" %
(frame_idx, 1, all_rewards[-1]))
losses_all.append(losses)
rewards_all.append(all_rewards)
mean_losses, var_losses = tl.StatShrink2D(losses_all)
mean_rewards, var_rewards = tl.StatShrink2D(rewards_all)
tl.save2D4list(mean_losses, var_losses, mean_rewards, var_rewards,
"data/noisydqn_CartPole_five1.npz")
replay_buffer = tl.BaseReplayBuffer(capacity)
update_target(current_model, target_model)
for frame_idx in range(1, num_frames + 1):
args_k = 0.
action = current_model.act(state)
next_state, reward, done, _ = env.step(action)
replay_buffer.push(state, action, reward, next_state, done)
state = next_state
episode_reward += reward
if done:
args_k = k_by_reward(episode_reward)
state = env.reset()
all_rewards.append(episode_reward)
episode_reward = 0
if len(replay_buffer) > replay_initial:
loss = improved_td_loss(batch_size, replay_buffer, current_model,
target_model, gamma, args_k, optimizer)
losses.append(loss.data[0])
if frame_idx % updatefrc == 0:
update_target(current_model, target_model)
if frame_idx % 2000 == 0:
print("This is %d frame of 1000000, k_end %d, reward is %f" %
(frame_idx, k_end, all_rewards[-1]))
save_dir = "data/analysis/rl1_kend%d" % (k_end * 2)
tl.save2D4list(losses, [], all_rewards, [], save_dir)
next_state, reward, done, _ = env.step(action)
replay_buffer.push(state, action, reward, next_state, done)
state = next_state
episode_reward += reward
if done:
args_k = k_by_reward(episode_reward)
state = env.reset()
all_rewards.append(episode_reward)
episode_reward = 0
if len(replay_buffer) > replay_initial:
loss = improved_td_loss(batch_size, replay_buffer, current_model,
target_model, gamma, args_k, optimizer)
losses.append(loss.data[0])
if frame_idx % updatefrc == 0:
update_target(current_model, target_model)
if frame_idx % 2000 == 0:
print("This is %d frame of 1000000, k_end %d, reward is %f" %
(frame_idx, k_end * 2, all_rewards[-1]))
save_dir = "data/analysis/rl05_kend%d" % (k_end * 2)
tl.save2D4list(losses, [], all_rewards, [], save_dir)
mean_losses, var_losses = tl.StatShrink2D(losses_all)
mean_rewards, var_rewards = tl.StatShrink2D(rewards_all)
#tl.save2D4list2(mean_losses, var_losses, mean_rewards, var_rewards, weight_sigmas, bais_sigmas, frame_list, save_dir)
tl.save2D4list(mean_losses, var_losses, mean_rewards, var_rewards, save_dir)