This letter investigates a sum rate maximizationproblem in an intelligent reflective surface (IRS) assisted non-orthogonal multiple access (NOMA) downlink network. Specif-ically, the sum rate of all the users is maximized by jointlyoptimizing the beams at the base station and the phase shiftat the IRS. The deep reinforcement learning (DRL), which hasachieved massive successes, is applied to solve this sum ratemaximization problem. In particular, an algorithm based on thedeep deterministic policy gradient (DDPG) is proposed. Both therandom channel case and the fixed channel case are studied inthis letter. The simulation result illustrates that the DDPG basedalgorithm has the competitive performance on both case.