Grant-Free Access with Multipacket Reception: Analysis and Reinforcement Learning Optimization

Grant-free access has been identified by 3GPP as a potential solution for Industrial Internet-of-Things applications in 5G networks. It allows to decrease overhead and delay, but it is also prone to collisions in the high-load regime. To reduce the effects of collisions, Non-Orthogonal Multiple Access or other Successive Interference Cancellation (SIC) protocols can be applied, allowing to partially recover collisions. In this paper, we abstract the grant-free access protocols with SIC with a $K$-Multipacket Reception ($K$-MPR) model. Based on this abstraction, we analyze its one-frame and steady-state throughput, delay and failure probability under different backoff schemes. Furthermore, we propose a reinforcement learning approach to allocate grant-free resources dynamically in order to maximize the normalized throughput of the protocol. Monte-Carlo simulations are employed to confirm the accuracy of analytical results and to evaluate the throughput, delay, and reliability of the proposed resource allocation approach.