env.reset()
done = False
curr_score = 0
for act_i in range(max_ep_len):
exp = agent.act(env, eps)
curr_score = curr_score + exp.reward
if exp.done:
break
if act_i % learn_every == 0:
agent.learn(64)
score_list.append(curr_score)
score_window.append(curr_score)
if episode % update_every == 0:
agent.update_target()
if episode % 20 == 0:
print("episode " + str(episode) + ", mean score: " +
str(np.mean(score_window)))
if episode % 100 == 0:
mean_score_list.append(np.mean(score_window))
print("test completed with scores: " + str(mean_score_list))
agent.save_checkpoint(local_checkpoint="test_out_3/qnet_" + env_sel +
"_local_test_" + str(test_i) + ".ckp")
pickle.dump((score_list, mean_score_list, pars),
open(
"test_out_3/qnet_" + env_sel + "_scores_and_pars_test_" +
str(test_i) + ".p", "wb"))
results.append(max(mean_score_list))
result_pars.append(pars)
score_window.append(curr_score)
#print and save
if episode % 20 == 0:
print("Episode " + str(episode) + ". Eps = " + str(eps) +
", mean_score: " + str(np.mean(score_window)))
ax.clear()
ax.plot(np.arange(len(score_list)), score_list)
plt.ylabel('Score')
plt.xlabel('Episode #')
plt.draw()
plt.pause(.001)
if np.mean(score_window) > env_pars.target_score:
agent.save_checkpoint(target_checkpoint="qnet_" + env_sel +
"_target_episode_" + str(episode) + ".ckp",
local_checkpoint="qnet_" + env_sel +
"_local_episode_" + str(episode) + ".ckp",
delayer_checkpoint="qnet_" + env_sel +
"_delayer_episode_" + str(episode) + ".ckp")
pickle.dump(
score_list,
open("qnet_" + env_sel + "_scores_" + str(episode) + ".p",
"wb"))
#final print and save
agent.save_checkpoint(target_checkpoint="qnet_" + env_sel +
"_target_final.ckp")
pickle.dump(score_list, open("qnet_" + env_sel + "_scores_final.p", "wb"))
plt.plot(np.arange(len(score_list)), score_list)
plt.ylabel('Score')
plt.xlabel('Episode #')
plt.draw()