ce_opt.py

# cross-entropy optimization
import numpy as np
import tqdm


def ce_minimize(X_mean, X_std, eval_loss, iterations=100, sample_size=25, keep_samples=10):
    assert 0 < sample_size
    assert 0 < keep_samples and keep_samples < sample_size
    # Alg settings:

    trange = tqdm.trange(iterations)
    for i in trange:
        # sample parameter vectors
        Xs = np.random.multivariate_normal(mean=X_mean,
                                           cov=np.diag(np.array(X_std**2)),
                                           size=sample_size)
        losses = np.array([eval_loss(x) for x in Xs])
        # keep the best samples
        keep_inds = losses.argsort()[:keep_samples]
        keep_Xs = Xs[keep_inds]

        X_mean = keep_Xs.mean(axis=0)
        X_std = keep_Xs.std(axis=0)
        trange.set_description("mean loss: %8.3g" % (np.mean(losses)))
    return X_mean


def test():
    def eval_loss(x):
        one_x, = x
        loss = (one_x - 5.)**2.
        return loss

    X_mean = np.array([0.])
    X_std = np.array([4.])

    X = ce_minimize(X_mean, X_std, eval_loss)
    np.testing.assert_array_almost_equal(np.array([5.]), X)
    print(X)