def lcb_f(x):
x = format_x(x, context)
mu, var = self.predict(x)
lcb = mu - scale * np.sqrt(var)
if negate:
lcb = -lcb # minimize
return lcb # maximize
def ac_f(x):
# TODO: acquisition predicate or unify with lcb_f
x = format_x(x, context)
mu, var = self.predict(x)
ret = mu / np.sqrt(var)
# self.beta < mu / np.sqrt(var)
# self.beta*np.sqrt(var) < mu
# 0 < self.beta*np.sqrt(var) - mu
return ret.T[0]
def lcb_fg(x):
x = format_x(x, context)
mu, var = self.predict(x)
f = mu - scale * np.sqrt(var)
dmdx, dvdx = self.predictive_gradients(x)
g = dmdx - 0.5 * dvdx / np.sqrt(var)
if negate:
f, g = -f, -g # minimize
return f[0, 0], g[0, self.func.param_idx] # maximize
def ac_f(x):
# x = helper.grid_around_point(x, self.robust_delta, n=3, x_range=self.func.x_range)
x = format_x(x, context)
mu, var = self.predict(x)
if self.use_var:
f = mu / np.sqrt(var)
else:
f = mu
if negate:
f = -f # minimize
return f # maximize
def ac_f(x):
x = format_x(x, context)
mu, var = self.predict(
x, noise=noise, **kwargs
) # self.model.predict_quantiles(x, quantiles=(2.5, 97.5))
# TODO: obtain gradient wrt query point for optimization
# if x.shape[0] == 1:
# #print(self.model.predict_jacobian(Xnew=x))
# dmu_dX, dv_dX = self.model.predictive_gradients(Xnew=x)
# dmu_dX = dmu_dX[:, :, 0]
# ds_dX = dv_dX / (2 * np.sqrt(var))
# #print(dmu_dX, ds_dX)
# #print(dmu_dX.shape, dv_dX.shape)
lse = -np.abs(mu) + scale * np.sqrt(var)
if negate:
lse = -lse # minimize
return lse # maximize
def ac_fg(x):
# TODO: could instead sample a fixed number of points
# x = helper.grid_around_point(x, self.robust_delta, n=3, x_range=self.func.x_range)
x = format_x(x, context)
mu, var = self.predict(x)
dmdx, dvdx = self.predictive_gradients(x)
if self.use_var:
f = mu / np.sqrt(var)
g = (np.sqrt(var) * dmdx -
0.5 * mu * dvdx / np.sqrt(var)) / var
# return np.average(f, axis=0)[0], np.average(g, axis=0)[self.func.param_idx]
g = g[0, :]
# return f[0, 0], g[0, self.func.param_idx]
else:
f, g = mu, dmdx
if negate:
f, g = -f, -g # minimize
return f[0, 0], g[self.func.param_idx] # maximize
def ac_f(x):
x = format_x(x, context)
_, var = self.predict(x, **kwargs)
if negate:
var = -var # minimize
return var # maximize
def ac_f(x):
x = format_x(x, context)
mu, _ = self.predict(x)
if negate:
mu = -mu
return mu