The proper calibration of the model values is crucial for high-precision ellipsometer measurement, and random error is critical in determining calibration accuracy. This study discusses random error estimation in the ellipsometer. The random errors encountered in measurement results are numerically evaluated for dark noise-limited, shot noise-limited, and light-source fluctuation-noise-limited systems. To reduce the random errors in measurement results, this study adopts the enumeration method and a genetic algorithm to find the best configuration for the aforementioned three noise-limited systems. Numerical analysis and experimental results show that compared with the commonly used configuration, the random error under the proposed optimal design can be reduced by more than 1/3. Although only the optimal configurations of three noise constrained systems are given in this paper, however, the noise estimation results and optimization methods can be employed in real-world models with various noise models.