print("MAE on validation data:", MAE)
MSE_valid = np.sum((sample_validation.rings - prediccions)**2) / N_valid
print("MSE validation MLP:", MSE_valid)
NMSE_val = sum((sample_validation.rings - prediccions)**2) / (
(N_valid - 1) * np.var(sample_validation.rings))
print("NMSE validation MLP:", NMSE_val)
R_squared = (1 - NMSE_val) * 100
print("Our model explain the {}% of the validation variance".format(R_squared))
#%% MLP MultiLayer
model_nnet = MLPRegressor(hidden_layer_sizes=[128, 64, 32],
alpha=0,
activation="logistic",
learning_rate='constant',
solver='lbfgs')
model_nnet.learning_rate_init = 1e-3
model_nnet.max_iter = 500
model_nnet.fit(sample.loc[:, "male":"shell_weight"], sample.rings)
print("Final loss 1st training module: ", model_nnet.loss_)
prediccions = model_nnet.predict(sample_validation.loc[:,
"male":"shell_weight"])
MAE = np.sum(abs(sample_validation.rings - prediccions)) / N_valid
print("MAE on validation data before refining:", MAE)
NMSE_val = sum((sample_validation.rings - prediccions)**2) / (
(N_valid - 1) * np.var(sample_validation.rings))
print("NMSE validation MLP before refining:", NMSE_val)
model_nnet.learning_rate_init = 1e-5
model_nnet.max_iter = 500
model_nnet.fit(sample.loc[:, "male":"shell_weight"], sample.rings)
print(model_nnet.get_params(), file=open('coeficients/mlp_multilayer', 'w'))
#print("Coeficients:", model_nnet.coefs_, "Biasis:", model_nnet.intercepts_)
def make_fig3():
names = ["iterations", "rewards"]
# df20k = pd.read_csv("tmp/lunar_lander_01/lunar_lander_20k.txt", names=names)
df1 = pd.read_csv("tmp/lunar_lander_03/lunar_lander_5k.txt", names=names)
df2 = pd.read_csv("tmp/lunar_lander_02_epochs=10/lunar_lander_5k.txt",
names=names)
avg_rewards_1 = pd.rolling_mean(df1["rewards"], 100)
avg_rewards_2 = pd.rolling_mean(df2["rewards"], 100)
fig1 = pl.figure(figsize=(12, 12))
ax1 = fig1.add_subplot(221)
ax1.plot(df1["iterations"], avg_rewards_1, "-", color="blue", lw=2)
ax1.plot(df2["iterations"], avg_rewards_2, "-", color="green", lw=2)
ax1.set_xlabel("episode", size=18)
ax1.set_ylabel("average reward", size=18)
ax1.text(2500, 220, "training epochs = 2", color="blue")
ax1.text(2500, 10, "training epochs = 10", color="green")
ax1.text(250, -175, "(a)", fontsize=22)
ax1.tick_params(labelsize=16)
np.random.seed(1)
ssize = 256
ones = np.ones((ssize, 1))
x = np.linspace(-1, 1, ssize).reshape((ssize, 1))
y = -0.5 + 2 * np.sin(2 * np.pi * x) + np.random.random((ssize, 1))
X = np.hstack((ones, x))
regressor = MLPRegressor(hidden_layer_sizes=(50, 50),
activation="relu",
batch_size=128,
max_iter=25,
solver="adam")
alpha_list = np.logspace(-6, -0.8, 1000)
loss = np.zeros(alpha_list.shape)
for idx, alpha_t in enumerate(alpha_list):
regressor.learning_rate_init = alpha_t
regressor.fit(X, y)
loss[idx] = regressor.loss_
print "learning rate = %0.6f, loss_ = %0.3f" % (alpha_t, loss[idx])
if np.abs(alpha_t - 1e-2) <= 1e-3:
yp = regressor.predict(X)
ax2 = fig1.add_subplot(222)
ax2.semilogx(alpha_list, loss)
ax2.set_xlabel("learning rate", size=18)
ax2.set_ylabel("fitting loss", size=18)
ax2.tick_params(labelsize=16)
ax2.text(2e-6, 0.1, "(b)", fontsize=22)
iax2 = inset_axes(ax2, width="30%", height=1., loc=1)
iax2.plot(x, y, "o", color="gray", mec="none", ms=2)
iax2.plot(x, yp, "-", color="red", lw=2)
iax2.set_xlabel("x", size=11)
iax2.set_ylabel("y", size=11)
iax2.tick_params(labelsize=9)
ax3 = fig1.add_subplot(223)
df2 = pd.read_csv(
"tmp/lunar_lander_04_epsilon_scheduled/lunar_lander_5k.txt",
names=names)
ax3.plot(df1["iterations"],
avg_rewards_1,
"-",
color="blue",
alpha=1,
lw=2)
ax3.plot(df2["iterations"],
pd.rolling_mean(df2["rewards"], 100),
color="green",
alpha=1,
lw=2)
ax3.set_xlabel("episode", size=18)
ax3.set_ylabel("average reward", size=18)
ax3.tick_params(labelsize=16)
ax3.text(250, -250, "(c)", fontsize=22)
ax3.legend([r"Decaying $\epsilon$", r"Adaptive $\epsilon$"], loc=2)
iax3 = inset_axes(ax3, width="50%", height=1., loc=4, borderpad=2.5)
iax3.plot([0, 50, 50, 100, 100, 150, 150, 200, 200, 500],
[0.1, 0.1, 0.05, 0.05, 0.02, 0.02, 0.01, 0.01, 0.005, 0.005],
"-",
color="gray",
lw=2)
iax3.set_xlim([0, 250])
iax3.set_xlabel("average reward", size=11)
iax3.set_ylabel(r"$\epsilon$ schedule", size=11)
iax3.tick_params(labelsize=9)
ax4 = fig1.add_subplot(224)
df1 = pd.read_csv(
"tmp/lunar_lander_04_epsilon_scheduled/lunar_lander_5k.txt",
names=names)
df2 = pd.read_csv("tmp/lunar_lander_05_gamma=0.75/lunar_lander_5k.txt",
names=names)
df3 = pd.read_csv("tmp/lunar_lander_06_gamma=1.00/lunar_lander_5k.txt",
names=names)
df4 = pd.read_csv("tmp/lunar_lander_07_gamma=0.995/lunar_lander_5k.txt",
names=names)
ax4.plot(df1["iterations"][:2500],
pd.rolling_mean(df1["rewards"][:2500], 100),
color="black",
alpha=0.9,
lw=2)
ax4.plot(df4["iterations"],
pd.rolling_mean(df4["rewards"], 100),
color="gray",
alpha=0.9,
lw=2)
ax4.plot(df2["iterations"],
pd.rolling_mean(df2["rewards"], 100),
color="green",
alpha=0.9,
lw=2)
ax4.plot(df3["iterations"],
pd.rolling_mean(df3["rewards"], 100),
color="magenta",
alpha=0.9,
lw=2)
ax4.set_xlabel("episode", size=18)
ax4.tick_params(labelsize=16)
ax4.legend([
r"$\gamma=0.990$", r"$\gamma=0.995$", r"$\gamma=0.750$",
r"$\gamma=1.000$"
],
loc=4)
ax4.text(125, -900, "(d)", fontsize=22)
fig1.savefig("fig0.pdf", bbox_inches="tight")
pl.close("all")
