Resource scheduling plays a critical role in managing congestion in modern networks. This problem is inherently complex, as it requires timely decision-making where previous allocations influence future resource availability. Traditional scheduling policies aim to guarantee Quality-of-Service (QoS) while ensuring fair resource distribution. More recently, Machine Learning (ML)-based schedulers have been proposed to enhance scheduling efficiency. However, existing solutions fail to explicitly integrate Guaranteed Flow Bit Rate (GFBR) constraints into Reinforcement Learning (RL)-based scheduling decisions. In this paper, we propose Delay-Responsive Intelligent Learning for Latency-sensitive QoS (DRILL-Q), a novel RL based, QoS-aware Medium Access Control (MAC) scheduler designed to manage multi-flow traffic while explicitly ensuring GFBR compliance. DRILL-Q leverages a reward function that incorporates Packet Delay Budget (PDB), Head-of-Line (HOL) delay, and priority levels to optimize scheduling decisions. Simulation results demonstrate that DRILL-Q successfully prioritizes traffic based on its QoS requirements, while significantly improving Jain’s fairness among flows with the same 5G QoS Identifier (5QI) values. Compared to traditional and deterministic scheduling algorithms, DRILL-Q achieves superior resource allocation, reducing delay and jitter while maintaining higher throughput under congested network conditions.

This paper presents an in-depth analysis of the Listen-Before-Talk (LBT) procedures required for cellular technologies operating in the unlicensed spectrum. We specifically investigate the 5G New Radio Unlicensed Standalone (NR-U SA) design under the most recent 3GPP standard-compliant Type 1 Channel Access Procedure (CAP). We highlight and address key challenges in adapting the asynchronous Type 1 CAP (i.e., load-based CAP) to NR-U’s synchronous slot-based framework and propose the Scheduled Type 1 CAP, which introduces a novel scheduling mechanism and implementation of additional sensing to facilitate seamless integration. We carry out our performance evaluation using a full-stack end-to-end network simulator and construct a realistic 3GPP-compliant coexistence scenario in which 5G NR-U SA and Wi-Fi networks operate indoors in unlicensed sub-7 GHz bands. The users of the 5G NR-U SA and Wi-Fi networks demand eXtended Reality applications. Through an extensive evaluation study, we analyse the interplay between the NR-U numerologies and the CAPs used in shared channel access and evaluate their impact on the end-to-end performance of both technologies by considering various quality of service metrics. The results reveal some of the main pitfalls of the LBT procedure defined by 3GPP in TS 37.213 and demonstrate the clear advantages of the proposed Scheduled Type 1 CAP. Our work provides valuable insights into CAPs for shared spectrum, proposes standard-compliant solutions, and lays the groundwork for future research by introducing a pioneering open-source network simulation platform for evaluating NR-U SA coexistence scenarios with the Scheduled Type 1 CAP.

5G networks are designed to support a variety of services with highly demanding Quality-of-Service (QoS) requirements. This opened the door for novel extended reality (XR) media applications to emerge with 5G. However, recent 5G field tests and system-level simulation studies show that further XR enhancements are required to support a massive adoption of XR services in 5G networks. Such enhancements are expected to come into play with 5G-Advanced. In this line, we propose and study an XR loopback mechanism that adapts the XR traffic to the instantaneous 5G network conditions by exploiting an XR application feedback. We propose various XR loopback algorithms, strategies, and parameters’ configurations and study their impact on the 5G end-to-end performance. We conduct extensive simulation campaigns by building realistic end-to-end 5G network scenarios with 3GPP mixed XR traffic setups. Results show that the proposed XR loopback mechanism can boost XR performance in 5G networks by adapting to 5G network conditions, while keeping the XR QoS requirements under control. We provide various insights and practical directions on XR loopback design that allow us to take full advantage of the 5G network capabilities and progress toward 5G-Advanced network design.

One of the main design targets of New Radio (NR) is to support multiple applications, including low-latency data transmissions. To achieve that, multiple features have been introduced, among which dynamic scheduling timings (denoted by K0, K1, K2) is the one which determines the delay between the different paired control and data transmissions. For example, K2 is the delay (in unit of slots) between an uplink grant reception and the corresponding uplink data transmission. The End-to-End (E2E) latency would be highly impacted by these scheduling timings. Based on the common observation, lower scheduling timings would lead to low E2E latency. However, in this paper, especially for uplink traffic, we show that this is not always true, and there are cases in which increasing the scheduling timings provides better E2E delay performance. This is due to the interplay between traffic patterns, gNB-UE process, and the NR numerology. To show such impact of scheduling timings on E2E latency with different uplink traffic patterns and NR numerologies, we implement an E2E ns-3 based NR simulator.

The 5G New Radio (NR) access technology defines multiple numerologies to support a wide range of carrier frequencies, deployment scenarios, and variety of use cases. In this paper, we consider a resource allocation problem to efficiently support multiple numerologies simultaneously. We assume frequency division multiplexing (FDM) of numerologies in a time division duplex (TDD) system with a self-contained slot format. We focus on optimizing the numerology subband (SB) configuration, as well as the duplexing ratio between downlink (DL) and uplink (UL) directions within each SB. The optimization problem minimizes the weighted sum of the normalized load (NL) for each SB in each direction. We prove that our optimization problem is convex and, furthermore, we derive the optimal closed-form expressions for the numerology SB configuration and the DL-UL duplexing ratio per SB. The effectiveness of the proposed resource allocation is validated through an end-to-end ns-3 based simulator, which shows how the optimization of the NLs is translated into an improved throughput and delay performance.

License Assisted Access (LAA) is the technology introduced by the Third Generation Partnership Project (3GPP) that enables the deployment of LTE networks in the unlicensed 5 GHz spectrum. To ensure a fair coexistence of LAA in the unlicensed spectrum with other technologies, e.g., with Wi-Fi, 3GPP has standardized the use of Listen Before Talk (LBT) as the default channel-access scheme for LAA. However, the performance of Wi-Fi when coexisting with LAA mainly relies on how the LBT parameters are configured by the LAA. In this paper, we focus on the Contention Window (CW) size parameter of LBT in LAA. We propose a Neural Network (NN) based scheme that adapts the CW size based on the predicted number of Negative Acknowledgments (NACKs) for all the subframes in a Transmit Opportunity (TXOP) of LAA. In particular, our proposed scheme learns from the past experience how many NACKs per subframe of a TXOP were received under certain channel conditions. The performance evaluation shows that our proposed scheme, when compared to the state-of-the-art approaches, provides the best trade-off between the fairness to Wi-Fi and the LAA performance in terms of both throughput and latency.

Cognitive networking paradigms may help meet the challenges of operating complex wireless communications networks. In this paper, we contrast the neural network (NN) and the Bayesian network (BN) models to extract information from real-time observations and optimize network performance. In particular, we apply these two models to the problem of call admission control (CAC) for a long term evolution (LTE) system. We simulate a realistic LTE scenario with mobility in ns-3 and we select the most relevant features that can be observed by the base station. Then, we design two new CAC schemes that autonomously learn the network behavior from the observation of the selected features. Furthermore, we propose a performance comparison among these two schemes and a state-of-the-art CAC scheme, showing that the NN and the BN schemes are very promising solutions for CAC in LTE systems.

Nowadays IEEE 802.11 WLANs are widely deployed; in spite of this, the issue of designing an efficient and practical Access Point selection schemes that can provide the best throughput performance in a variety of link conditions is still open. In this paper we present a Cognitive AP selection scheme that allows the mobile station to learn from its past experience how to select the best AP. In our proposal the mobile station collects measurements regarding the past link conditions and throughput performance, and a cognitive engine based on a Neural Network trained on this data drives the AP selection process. Our performance evaluation shows that the proposed scheme has very good performance in a variety of scenarios, as opposed to other algorithms previously proposed in the literature which perform well only in specific cases and cannot address the non-idealities typical under real conditions.