def __init__(self, x0, sigma0, n_pop, n_select, min_sigma=1-5, big_sigma=1e-1, wait_iter=10, min_incr=1e-2):

self.x = np.array(x0, dtype=np.float)

self.sigma = np.full_like(x0, sigma0, dtype=np.float)

self.min_sigma = min_sigma

self.big_sigma = big_sigma

self.n_pop = n_pop

self.n_select = n_select

self.waiting = False

self.gen = 0

self.best_score = -float('inf')

self.best_iter = 0

self.iter = 0

self.wait_iter = wait_iter

self.min_incr = min_incr

def ask(self):

if self.waiting:

raise Exception("A population has already been asked for")

self.waiting = True

self._pop = np.random.randn(self.n_pop, *self.x.shape) * self.sigma + self.x

self.gen += 1

return self._pop

def tell(self, scores):

if not self.waiting:

raise Exception("A population must be asked for")

self.iter += 1

self.waiting = False

inds_f = np.argsort(scores)[::-1]

parents = self._pop[inds_f[:self.n_select]]

min_sigma = self.min_sigma

best = scores[inds_f[0]]

if best - self.best_score > self.min_incr:

self.best_score = best

self.best_iter = self.iter

else:

if self.iter - self.best_iter > self.wait_iter:

min_sigma = self.big_sigma

self.best_iter = self.iter

print("Taking too long, varying weights")

print("Generation {}: best - {:0.6f}, elite - {:0.6f}, avg - {:0.6f}, worst - {:0.6f}".format(self.gen, scores[inds_f[0]], np.mean([scores[i] for i in inds_f[:self.n_select]]), np.mean(scores), scores[inds_f[-1]]))

self.sigma[:] = np.sqrt(np.mean(np.square(parents - self.x), axis=0)) + min_sigma