Non-orthogonal multiple access (NOMA) exploits the potential of the power domain to enhance the connectivity for the Internet of Things (IoT). Due to time-varying communication channels, dynamic user clustering is a promising method to increase the throughput of NOMA-IoT networks. This paper develops an intelligent resource allocation scheme for uplink NOMA-IoT communications. To maximise the average performance of sum rates, this work designs an efficient optimization approach based on two reinforcement learning algorithms, namely deep reinforcement learning (DRL) and SARSA-learning. For light traffic, SARSA-learning is used to explore the safest resource allocation policy with low cost. For heavy traffic, DRL is used to handle traffic-introduced huge variables. With the aid of the considered approach, this work addresses two main problems of fair resource allocation in NOMA techniques: 1) allocating users dynamically and 2) balancing resource blocks and network traffic. We analytically demonstrate that the rate of convergence is inversely proportional to network sizes. Numerical results show that: 1) Compared with the optimal benchmark scheme, the proposed DRL and SARSA-learning algorithms have lower complexity with acceptable accuracy and 2) NOMA-enabled IoT networks outperform the conventional orthogonal multiple access based IoT networks in terms of system throughput.