def search(self):
"""Search for an optimal solution."""
self.reset()
if not self.ready():
raise Exception("Optimizer not ready.")
best = GbSolution(None, float('-Inf'))
top_ranked = None
while not self.done():
self.iters += 1
candidates = self.sample(self.sample_size, top_ranked, best)
top_ranked, winner = self.get_top_ranked(candidates)
self.estimate(top_ranked, best)
if best.cost < winner.cost:
best = winner
if self.callback_func is not None:
self.callback_func(
best, self.cost_func.evals / float(self.max_evals))
if self.cost_func.retest_last:
best.cost = self.cost_func(np.array([best.params]),
is_last=True)[0]
if self.callback_func is not None:
self.callback_func(best, 1.0)
return best
def get_top_ranked(self, candidates):
costs = self.cost_func(candidates)
maxindx = np.argmax(costs)
winner, loser = GbSolution(candidates[maxindx],
costs[maxindx]), GbSolution(
candidates[not maxindx],
costs[not maxindx])
return winner, winner
def test_calculate_means_vars_no_best(self):
means = np.array([2.0])
vars = np.array([9.0])
acc_means = np.array([1.5])
acc_vars = np.array([4.0])
cand_size = 5
learning_rate = 0.6
best = GbSolution(None, float("-Inf"))
new_means, new_vars = \
Tilda._estimate_gaussian(means, vars, acc_means, acc_vars, best, cand_size, learning_rate)
self.assertAlmostEqual(new_means[0], 1.16, 2)
self.assertAlmostEqual(new_vars[0], 4.2681, 2)
def build_solution(self, params, cost):
return GbSolution(params, cost)
def build_solution(self, params, cost):
return GbSolution(params, cost, self.distr.roots, self.distr.children)