n_steps = 0
for i in range(n_games):
observation = env.reset()
done = False
score = 0
while not done:
env.render()
action, prob, val = agent.choose_actions(observation)
observation_, reward, done, info = env.step(action)
n_steps += 1
score += reward
agent.remember(observation, action, prob, val, reward, done)
if n_steps % N == 0:
agent.learn()
learn_iters += 1
observation = observation_
score_history.append(score)
avg_score = np.mean(score_history[-100:])
if avg_score > best_score:
best_score = avg_score
agent.save_models()
print(
f"episode {i}: score {score:.1f} avg score {avg_score: .1f} time_steps {n_steps} learning_steps {learn_iters}"
)
x = [i + 1 for i in range(len(score_history))]
plot_learning_curve(x, score_history, figure_file)
# have_grad = True
# if have_grad:
# observation_ = fgsm_attack(observation_, perturbation, data_grad)
data_grad = advAgent.compute_grads()
if data_grad is not False:
observation_ = fgsm_attack(observation_, perturbation, data_grad)
n_steps += 1
score += reward
agent.remember(observation, action, prob, val, reward, done)
advAgent.remember(observation, action, reward, observation_, done)
if not load_checkpoint:
if n_steps % N == 0:
loss.append(agent.learn())
learn_iters += 1
observation = observation_
if not load_checkpoint:
avg_loss = np.mean(loss, axis=0)
actor_loss.append(avg_loss[0])
critic_loss.append(avg_loss[1])
total_loss.append(avg_loss[2])
score_history.append(score)
avg_score = np.mean(score_history[-100:])
avg_score_history.append(avg_score)
score_book[trial] = score_history