Reinforcement Learning presents a promising method for enhancing neural machine translation models by utilizing policy optimization to maximize metric-based rewards. This paper presents a two-level neural machine translation method that integrates supervised learning and reinforcement learning for neural machine translation tasks. At the first level, we train the neural machine translation models using supervised learning to establish a strong foundational translation capability. In the second level, we employ reinforcement learning, using beam search for translation and BLEU scores as rewards to generate improved-quality translations. We carry out an empirical study on the English-Khasi language pair, and the results indicate that this two-level neural machine translation method significantly enhances the performance of neural machine translation.