def max(self, x):
s = x.dot(self.w)
# log_p = log_softmax(s)
# u = np.random.uniform(0.0, 1.0, s.shape)
# r = np.log(-np.log(u)) - log_p
# return argmax(-r)
return argmax(s)
def _draw_thompson(self, x):
# self._update_cholesky()
# v = 0.5 * np.sqrt(9 * self.d * np.log(max(1, self.t) / self.delta))
# u = np.random.standard_normal(self.w.shape)
means = np.dot(x, self.w)
stds = self.alpha * self._bound(x)
s = np.empty(self.k)
for i in range(self.k):
s[i] = np.random.normal(means[i], stds[i])
return argmax(s)
def max(self, x):
s = x.dot(self.w)
return argmax(s)
def draw(self, x):
s = x.dot(self.w)
log_p = log_softmax(s / self.tau)
u = np.random.uniform(0.0, 1.0, s.shape)
r = np.log(-np.log(u)) - log_p
return argmax(-r)
def max(self, x):
return argmax(x.dot(self.w))
def _draw_ucb(self, x):
s = np.dot(x, self.w) + self.alpha * self._bound(x)
s = np.minimum(1.0, s)
return argmax(s)