lrcall.py 文件源码

def get_next_by_EI(ni, alpha, lr, lr_time, X, y, ei_xi):
    '''
    Args:
        ni: number of units in the each layer
        alpha: lambda for Ridge regression
        lr: fitted performance model in burning period
        lr_time: fitted time model in burning period
        X: all previous inputs x
        y: all previous observations corresponding to X
        ei_xi: parameter for EI exploitation-exploration trade-off

    Returns:
        x_next: a nested list [[0,1,0], [1,0,0,0], ...] as the next input x to run a specified pipeline
    '''
    var = np.var(lr.predict(X) - y)
    m = np.dot(X.T, X)
    inv = np.linalg.inv(m + alpha * np.eye(sum(ni)))
    maxEI = float('-inf')
    x_next = None
    for i in range(np.prod(ni)):
        x = [[0]*n for n in ni]
        x_flat = []
        pipeline = get_pipeline_by_flatten_index(ni, i)
        for layer in range(len(ni)):
            x[layer][pipeline[layer]] = 1
            x_flat += x[layer]
        x_flat = np.array(x_flat)
        mu_x = lr.predict([x_flat])
        var_x = var * (1 + np.dot(np.dot(x_flat, inv), x_flat.T))
        sigma_x = np.sqrt(var_x)
        u = (np.min(y) - ei_xi - mu_x) / sigma_x
        EI = sigma_x * (u*norm.cdf(u) + norm.pdf(u))
        estimated_time = lr_time.predict([x_flat])[0]
        EIPS = EI / estimated_time
        if EIPS > maxEI:
            maxEI = EIPS
            x_next = x

    return x_next