The article discusses a new method called predecessor combination search (PCS) that improves the accuracy of neural networks by calibrating their output probabilities. This method helps prevent errors in critical applications by finding the best combination of earlier network blocks. PCS outperforms other techniques in calibration and enhances the network's ability to handle changes in data distribution. The research shows that early stopping, a common method to prevent overfitting, does not effectively calibrate neural networks. The results demonstrate that PCS achieves state-of-the-art performance in calibration and enhances the network's robustness.