Irregular Repetition Slotted Aloha (IRSA) is one candidate member of a family of random access-based protocols to solve massive connectivity problem for Internet of Things (IoT) networks. The key features of this protocol is to allow users to repeat their packets multiple times in the same frame and use Successive Interference Cancellation (SIC) to decode collided packets at the receiver. Although, the plain IRSA scheme can asympotically reach the optimal 1 [packet/slot]. But there are still many obstacles to achieve this performance, specially when considering short frame length. In this report, we study two new variants of IRSA with short frame length, and we optimize their performance using a Deep Reinforcement Learning approach. In our first variant, Random Codeword Selection-IRSA (RC-IRSA), we consider an IRSA approach with random codeword selection, where each codeword represents the transmission strategy of a user on the slots. We apply a Deep Reinforcement Learning to optimize RC-IRSA: we train a Deep Neural Network model that choses the slots on which the user sends its packets. Our DRL approach for RC-IRSA is a new optimization method for IRSA using a DRL approach and it works as a base for our second proposed IRSA variant DS-IRSA. Our second variant is a sensing protocol based on IRSA and trained with machine learning to synchronize the nodes during the transmission and avoid collisions. For that aim, we proposed DS-IRSA, Deep Learning Sensing-based IRSA protocol which is composed of two phases: a sensing phase, where the nodes can sense the channel and send short jamming signals, followed by a classical IRSA transmission phase. We use a DRL algorithm to optimize its performance. Our proposed protocol has shown an excellent performance to achieve an optimal performance of almost 1 [decoded user/slot] for small frame sizes (≤ 5) slots and with enough sensing duration.