Semi-decentralized Message Allocation for Cell-Free Networks using Unsupervised Learning

In this paper, we present an unsupervised learning framework for message allocation in Cell-free networks (CFNs) for latency minimization. One of the key features of CFNs is that users' data can be decoded by multiple access points (APs), i.e., in a "cell-free" manner by letting users connect to multiple APs simultaneously; leading to the problem of message allocation across APs. While a fully centralized approach for message allocation can make the most out of the flexibility offered via CFNs, it requires prohibitive coordination overhead. In this paper, we instead propose a novel semi-decentralized machine learning based framework for message allocation. It allows each user to split their messages using a "learned" model (e.g., a neural network) which takes two inputs: a user's local channel gains and aggregate global SINRs at the APs. The model is trained with the objective of minimizing the latency of the network. To accomplish this, the total latency derived from the model's learned message split for each user is the main component of the loss used to update the model. Different methods are investigated for training the model by presenting variations of how the latency is computed. We use the cumulative distribution function (CDF) of latency as the key performance metric and compare our proposed semi-decentralized approach against several centralized methods as well as uniform and greedy message allocation techniques. Our results indicate that the semi-decentralized machine learning based method can approach the performance of the centralized methods with very little coordination overhead and outperforms greedy/uniform allocation methods.