optimizer.py 文件源码

def _get_eis(self, points, score_optimum):
        """
        Calculate the expected improvements for all points.

        Parameters
        ----------
        points: list
            List of parameter settings for the GP to predict on

        score_optimum: float
            The score optimum value to use for calculating the difference against the expected value

        Returns
        -------
        Returns the Expected Improvement
        """

        # Predict mu's and sigmas for each point
        mu, sigma = self.score_regressor.predict(points, return_std=True)

        # Subtract each item in list by score_optimum
        # We subtract 0.01 because http://haikufactory.com/files/bayopt.pdf
        # (2.3.2 Exploration-exploitation trade-of)
        diff = mu - (score_optimum - 0.01)

        # Divide each diff by each sigma
        Z = diff / sigma

        # Calculate EI's
        ei = diff * norm.cdf(Z) + sigma * norm.pdf(Z)

        # Make EI zero when sigma is zero (but then -1 when sigma <= 1e-05 to be more sure that everything goes well)
        for index, value in enumerate(sigma):
            if value <= 1e-05:
                ei[index] = -1

        return ei