Non-orthogonal multiple access(NOMA) can be considered a candidate for next-generation communications networks due to its high spectral efficiency and ability to reduce latency by superposing signals of users in a user-pair at the same time and same frequency. NOMA has a key feature that the near user(NU) has prior information about the messages of the far user(FU) in the same user pair. It means that NU can be used for retransmission when the FU requests retransmission. However, retransmission from the NU can cause high power consumption which is not suitable for mobile stations or IoT devices. Therefore, adaptive power allocation is studied to reduce power consumption in this paper. Deep reinforcement learning(DRL) is used to decide the power for retransmission from the NU. Simulation results showed that the proposed DRL scheme can reduce power allocation and improve better reliability over the conventional retransmission scheme.