Toward Efficient Gradient-Based Value Estimation