agent = DQNAgent(config)
agent.create_dirs()
eps = config.eps_start
for i_episode in range(1, config.n_episodes + 1):
# Reset the environment and the score
env_info = env.reset(train_mode=True)[brain_name]
state = env_info.vector_observations[0]
score = 0
while True:
action = agent.act(state, eps)
env_info = env.step(action)[brain_name]
next_state, reward, done = env_info.vector_observations[
0], env_info.rewards[0], env_info.local_done[0]
agent.step(state, action, reward, next_state, done)
state = next_state
score += reward
if done:
break
scores_window.append(score)
scores.append(score)
avg_scores.append(np.mean(scores_window))
eps = max(config.eps_min, config.eps_decay * eps)
print(
'\rEpisode {}\tEps {:.2f}\tLast Score: {:.2f}\tAverage Score: {:.2f}'
.format(i_episode, eps, score, np.mean(scores_window)),
end="")
if i_episode % 100 == 0:
test_scores.append(test(env, agent, i_episode))
test_scores_i.append(i_episode)
env_info = env.reset(train_mode=True)[brain_name]
state_raw = process_state(env_info.visual_observations[0])
state_window.append(state_raw)
state = np.vstack(
[np.expand_dims(np.array(s), 0) for s in state_window])
score = 0
while True:
action = agent.act(state, eps)
env_info = env.step(action)[brain_name]
next_state_raw, reward, done = process_state(env_info.visual_observations[0]), env_info.rewards[
0], \
env_info.local_done[0]
state_window.append(next_state_raw)
next_state = np.vstack(
[np.expand_dims(np.array(s), 0) for s in state_window])
agent.step(np.array([state]), action, reward,
np.array([next_state]), done)
state = next_state
score += reward
if done:
break
scores_window.append(score)
scores.append(score)
avg_scores.append(np.mean(scores_window))
eps = max(config.eps_min, config.eps_decay * eps)
print(
'\rEpisode {}\tEps {:.2f}\tLast Score: {:.2f}\tAverage Score: {:.2f}'
.format(i_episode, eps, score, np.mean(scores_window)),
end="")
if i_episode % 100 == 0:
test_scores.append(test(env, agent, i_episode,
state_len=state_len))