Fingerprint recognition is a crucial biometric authentication technology, known for its uniqueness and stability. Despite recent advancements in fingerprint matching, fingerprint distortions still pose significant challenges to fingerprint matching algorithms. To address this, fingerprint registration techniques have been developed. While supervised dense registration shows promise, it is hindered by speed, accuracy, and data annotation challenges. Unsupervised methods offer a solution to the lack of labeled data to deal with incorrect matches and weak regularization. Inspired by reinforcement learning (RL), this paper introduces an RL-based approach for fingerprint registration, decomposing the process into manageable steps and leveraging the Normalized Cross-Correlation (NCC) reward for unsupervised learning. By decomposing the registration process into incremental steps, our method effectively balances the trade-off between accuracy and the need for extensive data annotation. Experimental results on the FVC2004 dataset demonstrate improved performance, adaptability, and reduced reliance on labeled data. This study enhances the efficiency of fingerprint registration, presenting a robust approach for handling complex deformation fields.