rw.py 文件源码

def NLS_annealing(F, xi, yi, p, N=100, n=10, sigma=5.,factor=0.5):
    # N = size of population in one iteration
    # n = number of iterations
    # sigma = initial (multiplicative) standard deviation
    # factor = factor to reduce sigma per iteration
    print "initial", p
    p = np.atleast_1d(p)
    dim = len(p)
    # make initial sigma act like multiplication by sigma^(+-1)
    sigma = np.log(sigma)*np.ones(dim)

    for k in range(n):
        # create new population by adding multiplicative gaussian noise
        P = p[None, :] * np.exp(np.random.randn(N, dim) * sigma[None, :])
        # compute mean square loss on population
        f = np.mean((F(xi[None, :], P) - yi)**2, 1)
        # replace p by new best guess
        p = P[np.argmin(f), :]
        # update sigma
        sigma *= factor
        print "parameters:", p
    print "minimum", min(f)

    return tuple(p)