The Internet of Things (IoT) aims to connect billions of devices, most of which are power and memory-constrained. Such constraints require efficient network access. "Irregular Repetition Slotted Aloha'" (IRSA) meets such requirements. In this paper, we optimize IRSA using Deep Reinforcement Learning to obtain Deep-IRSA, and introduce variants that allow re-transmission and user priority classes. We observe the learned degree distribution and throughput, showing that Deep-IRSA performs excellently, is generic, and could well replace known approaches for smaller frame sizes and IRSA variants.