thetas_archive = Archive(max_size=n_train)
# sampler
sampler = Samplers.BasicSampler(sample_archive, thetas_archive)
# training
for i in range(n_train):
batch = optimizer.ask(pop_size)
scores = np.zeros(pop_size)
# newly drawn samples
for j in range(pop_size):
nn.set_params(batch[j])
score = env.eval(nn, render=False)
sample_archive.add_sample(Sample(batch[j], score, [curr_gen]))
scores[j] = score
# statistics on the current batch
print("Best/Average score in pop {0}: {1:.2f}, {2:.2f}".format(
i, np.max(scores), np.mean(scores)))
# optimization step
optimizer.tell(batch, scores)
curr_gen += 1
# testing best
best_params = optimizer.best_param()
nn.set_params(best_params)
for i in range(n_test):
score = env.eval(nn, render=True)
env = Env("CartPole-v0")
act_space = env.get_action_space()
obs_space = env.get_obs_space()
# optimization stuff
nn = FFIndiv(obs_space, act_space, hidden_size=2)
best_params = nn.get_params()
best_score = env.eval(nn, render=False)
optimizer = ES.CMAES(nn.get_params().shape[0], pop_size=pop_size)
# archives
sample_archive = Archive(max_size=n_train * pop_size)
thetas_archive = Archive(max_size=n_train)
mu, cov = optimizer.get_distrib_params()
thetas_archive.add_sample(Theta(mu, cov, []))
# sampler
sampler = Samplers.ClosestSampler(sample_archive,
thetas_archive,
accept_ratio=accept_ratio)
print("Problem dimension:", mu.shape[0])
df = pd.DataFrame(columns=[
"n_reused", "best_score", "average_score", "sample_time",
"evaluation_time"
])
# training
for i in range(n_train):
