Multi-Connectivity in 5G and Beyond

• Multi-Connectivity (MC) is a wireless paradigm that maintains simultaneous connections to multiple access points or RATs, ensuring service continuity and enhanced reliability.
• MC architectures leverage a centralized unit (CU) and distributed units (DUs) to dynamically split, duplicate, and manage traffic based on real-time network conditions.
• AI-driven optimization, including reinforcement learning, is employed to adapt traffic flows and resource allocation, yielding significant improvements in throughput and latency.

Multi-Connectivity (MC) refers to a wireless networking paradigm in which user equipment (UE) maintains simultaneous connections to multiple access points, base stations, or radio access technologies (RATs) to achieve service continuity, enhance throughput, and boost reliability. MC is fundamental in heterogeneous and challenging environments such as 5G terrestrial-satellite integrated networks, millimeter-wave (mmWave) networks, and mission- or latency-critical network domains. MC’s defining feature is its native support for splitting or duplicating user traffic across spatially or technologically disjoint communication paths, leveraging diversity in link conditions and network resources (Lisi et al., 2020).

1. Architectural Frameworks and Protocol Organization

MC frameworks in advanced networks employ a clear architectural separation between centralized and distributed processing, typically by dividing the next-generation Node B (gNB) or network controller into a Centralized Unit (CU) and multiple Distributed Units (DUs). This separation facilitates technology-independent control and management in the CU (e.g., RRC, SDAP, PDCP, cRRM, traffic flow control), while the DUs focus on real-time, RAT-specific operations (RLC, MAC, PHY, local RRM, per-UE QoE estimation). The control-user plane protocol stack is typically split so that:

• The control plane (RRC) resides in the CU,
• Upper user plane layers (SDAP, PDCP) execute in the CU,
• Lower user plane layers (RLC, MAC, PHY) execute in the DUs.

User-plane traffic can be dynamically switched, duplicated, or split at the PDCP layer, enabling fast rerouting without costly control-plane reestablishment (Lisi et al., 2020). Dynamic session establishment leverages UE measurement reports, real-time monitoring, and global network resource knowledge at the CU to anchor MC sessions, select and configure secondary links, and coordinate traffic distribution across terrestrial and satellite DUs.

2. Optimization and Control Algorithms

Effective MC requires multidimensional optimization of throughput, reliability, latency, load, and overhead. The traffic flow control problem can be formulated as a constrained optimization in which, for active UE set $U$ and DU/RAT set $R$, continuous variables $x_{u,r} \in [0,1]$ encode the split/duplication ratio of each UE’s traffic. Objective functions are typically weighted sums: $$\max_{\{x_{u,r}\}} \sum_{u\in U} [ \alpha \cdot \sum_{r} R_{u,r}(x_{u,r}) - \beta \cdot \max_{r} L_{u,r}(x_{u,r}) + \gamma \cdot (1-\prod_{r}(1-p_{u,r}(x_{u,r})) ) ]$$ subject to per-DU capacity, conservation and feasibility constraints. Here $R_{u,r}, L_{u,r}, p_{u,r}$ model rate, latency, and reliability for the $u \rightarrow r$ path, while $(\alpha,\beta,\gamma)$ tune KPI priorities (Lisi et al., 2020).

This resource control is regularly recomputed by the CU through joint optimization or AI-based controllers and enforced via explicit scheduling and buffer-allocation commands to DUs. DUs locally enforce HARQ and short-timescale constraints, while the CU maintains global control over high-level traffic splitting.

3. AI-Driven MC: Reinforcement Learning Approaches

To cope with the complexity and dynamics of real-world MC—including the need for fast adaptation to link quality fluctuations, path failures, or user mobility—MC control problems are cast as Markov Decision Processes (MDPs) and solved via reinforcement learning (RL) algorithms.

The MDP state space encodes instantaneous SNRs, buffer levels, latency/reliability metrics, and QoS targets for all ($u$, $r$) pairs. Actions select DU subsets and traffic-splitting vectors per UE. Typical reward functions are weighted combinations of throughput, latency penalties, and reliability metrics: $$r = \sum_{u} [ w_1 \cdot \mathrm{Throughput}_u - w_2 \cdot \mathrm{Latency}_u + w_3 \cdot \mathrm{Reliability}_u ]$$ Q-learning or Deep Q-Networks (DQN) are used to learn control policies capable of both short-term adaptation (e.g., real-time AP reselection under blockage) and long-term reward maximization. The RL agent periodically queries network state, selects MC actions, and issues configuration updates to the DUs (Lisi et al., 2020).

4. Performance Metrics and Empirical Results

Performance of MC frameworks is evaluated along standard network KPIs:

• Throughput: average per-UE data rate ($R_u$),
• Latency: expected per-packet delay,
• Reliability: delivered/total packet fraction.

In the 5G-ALLSTAR framework, MC yields ≈30% throughput gain over single connectivity under mixed terrestrial/satellite load, up to 50% reduction in tail (95th-percentile) latency with packet duplication, and reliability improvement from ≈99% to ≈99.999% through PDCP-level duplication (Lisi et al., 2020). These gains reflect the significant practical benefits of MC’s link/path diversity and intelligent traffic management.

5. Operational Advantages, Challenges, and Extensions

Benefits:

• Seamless service continuity, even under terrestrial link outages or satellite blockages, via fast switching and session reconfiguration enabled by the PDCP split and CU-centric intelligence.
• Enhanced throughput from concurrent use or aggregation of heterogeneous RATs/links.
• Ultra-reliability via independent-branch replication and joint scheduling.
• Optimal resource usage stemming from a centralized, globally-informed control plane.

Challenges:

• Excessive signaling on the CU⇆DU interface if action rates are too high.
• Stringent requirements on UE-side buffering and reordering to handle path-delay disparities and packet duplication.
• Algorithmic complexity and scalability in very large-scale network deployments with many UEs and access points.

Ongoing Research Directions:

• Scalability of RL- and optimization-based MC control to scenarios with dense user populations and high access point counts.
• Cross-layer schemes integrating higher-layer feedback (e.g., transport-level congestion or application-layer QoS) into MC decision logic.
• Support for per-slice and multi-tenant MC in virtualized, partitioned networks.
• Exploitation of mobile edge computing to further decrease latency in MC control loops.
• Exploration of distributed, federated, or semi-supervised AI agents for workload sharing and privacy compliance across network domains (Lisi et al., 2020).

6. MC in Diverse Wireless Contexts

MC principles generalize across a spectrum of contemporary wireless scenarios:

• mmWave Networks: Centralized controllers use fast link switching and context preparation to counteract the rapid link intermittency of mmWave bands; MC yields 30–40% throughput gains and significantly lowers UE-level interruptions (Tatino et al., 2018).
• Learning-based Scheduling: MC scheduling can be nearly optimal via machine learning classifiers (e.g., random forests), using only alignment-phase SNR and minimal feedback, and outperforming SNR-based heuristics (Tatino et al., 2020).
• Indoor and THz Systems: MC’s benefit depends on fast adaptation (closest-instantaneous-LOS MC is superior to reactive MC under heavy fading/absorption) (Shafie et al., 2020).
• Mission- and Ultra-Reliability Scenarios: MC is indispensable for URLLC and mixed-criticality applications, with power/rate allocation and coding strategies co-designed to enable diversity gain without excessive resource penalties (Karacora et al., 2022).

7. Summary Table: Key MC Functions in 5G-ALLSTAR

Component	Role	Key Mechanisms / Layers
gNB-CU (central unit)	Non-real-time control, global optimization	RRC, SDAP, PDCP, Traffic Flow Control
gNB-DU (distributed)	Real-time, RAT-dependent processing	RLC, MAC, PHY, dRRM, QoE estimation
F1 interface	CU-DU signaling, config/coordination	Standardized, all MC session control
MC session management	Access point selection, traffic split/dup	QoS-based, AI-supported in CU
Traffic Flow Control	Throughput-latency-reliability optimization	Mathematically-formulated, AI-based
RL-based MC control	Dynamic DU selection, adaptation	Q-learning/DQN with network feedback
Metrics	Throughput, latency, reliability	Measurement & reconfiguration feedback

The 5G-ALLSTAR MC framework exemplifies the convergence of protocol-layer innovations, formal optimization, and AI-based resource allocation to deliver flexible, high-reliability, low-latency multi-path connectivity integrating terrestrial and satellite links, and is representative of state-of-the-art MC deployment (Lisi et al., 2020).