Random Forests Relay Selector in Buffer-Aided Cooperative Networks

This paper presents a data-driven Random Forest (RF) framework for the joint relay–link selection and buffer management problem in buffer-aided cooperative relay networks, with the goal of minimizing outage probability. Buffer-aided relaying exploits dynamic link selection to enhance throughput and reliability, but the underlying mixed-integer optimization is computationally intractable for even moderate network sizes. We recast relay selection as a multi-class classification task by generating “true” labels via a time window-based outage minimization procedure that captures long-term buffer–channel interactions. An RF classifier learns the mapping from system state features like channel gain, buffer length, link availability to optimal link labeling. A key benefit of the proposed framework is that the computationally intensive training phase is performed offline, and the resulting RF model can, in principle, be a strong candidate for real-time applications. Extensive simulations for networks with 3 and 6 relays and buffer sizes $\in $ [1, 50] demonstrate that the RF-based selector consistently achieves the lowest outage probability and outperforms the available conventional schemes especially in small-buffer networks, thereby making RF a promising tool for real-time cooperative network control.