In this paper, the energy efficiency of a two-user downlink non-orthogonal multiple access (NOMA) system is investigated. We consider a stochastic time-varying channel that is unknown at the transmitter side (in state and distribution) and propose an online learning method that is energy efficient and relies only on a 1-bit feedback, of high relevance in IoT networks. Our adaptive NOMA scheme draws on multi armed bandits (MAB) to learn the optimal policy in terms of which user should perform successive interference cancellation (SIC) and the power allocation at the transmitter. Our simulation results reveal fundamental tradeoffs between performance, complexity and available feedback information, and also show that our scheme can remarkably outperform OMA without CSIT/CDIT.