Goal: Various methods have been developed to analyze the association between organisms and their genomic sequences. Among them, sequence alignment is the most frequently used method for comparative analysis of biological genomes. We intend to propose a novel pairwise sequence alignment method using deep reinforcement learning to break out the old pairwise alignment algorithms. Methods: We defined the environment and agent to enable reinforcement learning in the sequence alignment system. This novel method, named DQNalign, can immediately determine the next direction by observing the subsequences within the moving window. Results: DQNalign shows superiority in the dissimilar sequence pairs that have low identity values. And theoretically, we confirm that DQNalign has a low dimension for the sequence length in view of the complexity. Conclusions: This research shows the application method of deep reinforcement learning to the sequence alignment system and how deep reinforcement learning can improve the conventional sequence alignment method.