Uplink Carrier Aggregation (UL-CA)

• Uplink Carrier Aggregation (UL-CA) is a technique that combines multiple uplink channels to enhance data transmission and network reliability.
• It is widely applied in cellular and multi-RAT networks to improve spectrum utilization and support high-capacity applications.
• Integrating UL-CA with dynamic traffic steering enables real-time resource allocation optimization and reduced latency.

Dynamic traffic steering policies are algorithmic frameworks designed for real-time adaptation of flow assignments or resource allocations across multiple available links (e.g., frequency bands, access points, RATs, or multi-hop routes) to optimize metrics such as throughput, delay, quality of service (QoS), and fairness. Unlike static policies, dynamic schemes leverage measurements of instantaneous load, congestion levels, demand estimates, or predictions to continuously adjust traffic allocation in response to time-varying network and traffic conditions. These policies are foundational for multi-link WLANs, cellular multi-RAT deployments, SDN-based networks, and energy-constrained or AI-driven wireless systems.

1. Principles and Taxonomy of Dynamic Traffic Steering

Dynamic traffic steering exploits real-time system observations to continuously direct packets, flows, or sessions toward links or paths with excess capacity or lower delay. Three principal approaches are commonly distinguished:

• Early Steering: Decision is made before link contention/arbitration, at the packet enqueue event. Policies may classify packets by flow or traffic class and select a link based on static mapping, current queue lengths, link utilization, or predicted delay and loss metrics. Early steering policies can be static (per-flow assignment) or dynamic (leveraging real-time per-link statistics) (Cena et al., 2024).
• Late Steering: Link selection occurs upon resource acquisition (e.g., after a TXOP is won), potentially enabling the controller to react to instantaneous channel, queue, and resource metrics that have changed since enqueue. Ideal late-steering may exploit up-to-the-microsecond state but is limited by firmware complexity on real hardware (Cena et al., 2024).
• Split/Combined Steering: Host-level software encodes, for each packet and for every retry (or transmission attempt), a bitmap of eligible links or bands, with link selection finalized at transmission. This grants full flexibility with minimal hardware/firmware complexity, and enables optimization against multiple objectives such as delay, throughput, and jitter (Cena et al., 2024).

In infrastructure WLANs such as IEEE 802.11be/ax/be (Wi-Fi 6/7), as well as SDN and SD-WAN overlays, dynamic traffic steering allows for effective load balancing across heterogeneous channels and mitigates single-link congestion, resulting in significant end-to-end performance gains (López-Raventós et al., 2022, Quang et al., 2023, Cena et al., 2024).

2. Formal Models and Optimization Frameworks

Dynamic steering policies are formalized using:

• Resource Allocation Optimization: Given link capacities $C_k$, loads $\ell_i$ per flow (or application), and measured occupancy $\kappa_k$ on each link, the allocation $x_{i,k}$ of flow $i$ over enabled interfaces $J_i$ solves:

$$\begin{aligned} &\max \sum_{i\in F} \sum_{k\in J_i} U(x_{i,k}) \ &\text{s.t.}~ \sum_{i\in F} x_{i,k} \leq C_k,~\forall k,~~\sum_{k\in J_i} x_{i,k} = \ell_i,~\forall i \end{aligned}$$

with $U(x)$ representing the system objective (e.g., sum throughput). Link-specific metrics (free airtime, predicted delay) drive allocation (López-Raventós et al., 2022).

• Markov Decision Processes (MDPs): For more complex and stochastic environments, steering policy design is cast as an MDP $$\mathcal{M} = (\mathcal{S}, \mathcal{A}, P, R, \gamma)$$. The state encodes recent system traffic, load, and link condition; the action is a mapping of flows to links/paths; the reward captures throughput, delay, load balance, or policy objectives. Deep RL agents (DQN, PPO, Actor-Critic) are widely used to optimize these policies under uncertainty and varying dynamics (Sun et al., 2023, Xu et al., 2023, Habib et al., 2023, Mushtaq et al., 2022).
• Hierarchical and Cascade RL: To manage high dimensionality, state decomposition and policy factorization are adopted, with separate agents specializing in subspaces or time scales, then combined through meta-controllers or soft classification (Sun et al., 2023, Habib et al., 2024).

3. Algorithmic Realizations and Policy Adaptation

Dynamic steering policy implementations share several technical ingredients:

• Measurement and Adaptation: At each adaptation epoch (e.g., new flow arrival or periodic timer), channel occupancy, queueing delay, throughput, and other KPIs are collected per link. Flows are assigned to links based on proportional free airtime, predicted speed, or estimated cost functions, with assignment updated as new measurements arrive (López-Raventós et al., 2022, Cena et al., 2024).
• Flow Prioritization and Specialization: Flows may be prioritized by application type (e.g., video vs. BE), and steering decisions are made to maximize link diversity and minimize flow starvation. For heterogeneous QoS classes, the allocation policy accommodates both the number of eligible interfaces and flow deadlines (López-Raventós et al., 2022).
• Reinforcement Learning/RL Policy Selection: In large, unpredictable environments, RL is used to learn steering policies that maximize composite KPIs. Techniques include:
  • Policy banks with daily similarity-based selection to adapt to previously unseen load scenarios, achieving near-oracle performance without retraining (Xu et al., 2023).
  • Federated and multi-agent DQNs for per-device steering under privacy/personalization or computational constraints (Zhang et al., 2023).
  • Hierarchical RL with meta-controller/goal-setting at slow time scale and controller-level, per-flow steering at fast time scale, shown to outperform heuristics and flat RL in O-RAN (Habib et al., 2024).
• Efficient Algorithms and Complexity Considerations: Implementations must limit $O(|F| \cdot |K|)$ complexity and ensure update intervals (e.g., $\delta=1$ s) are feasible for real-time adaptation in dense deployments (López-Raventós et al., 2022).

4. Performance Evaluation and Empirical Results

Dynamic traffic steering policies robustly outperform static or heuristic baselines across diverse metrics and network settings:

• IEEE 802.11be WLANs: In flow-level simulation with multi-link operation, the MCAB policy (periodic dynamic reallocation) achieves a 17% improvement in worst-case BSS-average satisfaction over the best non-dynamic policy, maintaining $\bar s \geq 0.95$ in more than 90% of scenarios. Interface occupancy is continuously balanced, avoiding persistent overloading seen with static assignment (López-Raventós et al., 2022).
• Wi-Fi 7 MLO: Early+Late per-packet steering, with per-packet and retry-based bitmaps, enables fine-grained real-time load balancing, expected to yield 20–30% throughput gains and 30–50% tail-latency reduction under realistic load heterogeneity (Cena et al., 2024).
• 5G Multi-RAT RL Steering: Deep Q-learning for dual-connectivity LTE/5G increases system throughput by 6–10% and reduces network delay by 23–33% compared to Q-learning and heuristic policies, dynamically steering traffic according to SINR, queue length, and per-flow QoS class (Habib et al., 2023).
• O-RAN and SDN/SD-WAN: Cascade RL and hierarchical DQN enable robust scaling and fast adaptation. For instance, in digital twin evaluations, CaRL improves cluster-aggregated downlink throughput by 24% and 18% (two different city clusters) over business-as-usual policies (Sun et al., 2023). In SD-WAN, cross-traffic-aware dynamic load balancing improves SLA satisfaction by up to 40% relative to static allocation (Quang et al., 2023).
• Intelligent Transportation Systems (ITS): In AV micro-simulations, DRL-based dynamic traffic steering reduces cumulative delay by up to 34% compared to fixed signaling, and rerouting decreases intersection queue lengths by 30% (Mushtaq et al., 2022).

5. Practical Considerations, Design Extensions, and Limitations

Practical deployment of dynamic traffic steering faces several challenges and consideration points:

• Measurement and Feedback Overhead: Steering decisions require real-time or near-real-time collection of per-link and per-flow metrics, which may stress control-plane bandwidth or hardware limits. Aggregation over windows or measurement sampling can balance precision and cost (Habib et al., 2023).
• Scalability and State Space Factorization: As state/action spaces grow (e.g., in O-RAN or SD-WAN), factorization and decomposition into smaller per-subspace policies or hierarchical meta-controllers mitigates sample inefficiency and overfitting (Sun et al., 2023, Habib et al., 2024).
• Granularity of Adaptation: While some policy-selection frameworks operate at the daily or hourly interval, finer timescales may be needed for highly dynamic environments or latency-critical applications (Xu et al., 2023).
• Policy Generalization and Robustness: RL-based methods can degrade when deployed in out-of-distribution scenarios not covered by trained policy banks; similarity metrics and online adaptation are used but true generalization remains challenging (Xu et al., 2023, Zhang et al., 2023).
• Integration with Control Architectures: ORAN-based traffic steering leverages clear separation of time scales and the use of non-RT (policy and prediction) versus near-RT (resource allocation) RICs, following standardized interfaces (A1, E2, etc.). Corresponding decomposition aligns with the natural control hierarchy of modern wireless systems (Kavehmadavani et al., 2022, Sun et al., 2023, Kavehmadavani et al., 2023).
• Multi-objective and Application-Aware Steering: Modern policies integrate objectives such as energy consumption (in green networking), delay/reliability (URLLC), or fairness, often via weighted reward combinations or multi-stage optimization hierarchies (Zhang et al., 2017, Kavehmadavani et al., 2023).
• Explainability and Policy Verification: Learned policies, especially those using neural approximators or implicit clustering, can be opaque. Classifier-based similarity measures in policy banks, and offline RL techniques with REM and CQL regularization, are employed to mitigate overestimation or instability, but interpretability continues to be an area under active research (Xu et al., 2023, Lacava et al., 2022).

6. Application Domains and Empirical Benchmarks

Dynamic traffic steering underpins a wide range of practical and emerging networking domains:

Domain/Stack	Policy Class	Empirical Gain	Citation
Wi-Fi 6/7 MLO	Per-packet per-retry, dynamic	17%–30% throughput↑, 30%–50% latency↓	(López-Raventós et al., 2022, Cena et al., 2024)
5G/ORAN xApp	RL, hierarchical, bank/selector	15%–24% throughput↑, 27%–59% delay↓	(Sun et al., 2023, Habib et al., 2024)
Multi-RAT 5G	DQN traffic/RAT assign.	6%–10% throughput↑, 23%–33% delay↓	(Habib et al., 2023)
SD-WAN Overlay	Cross-traffic dynamic alloc.	Up to 40% SLA satisfaction↑	(Quang et al., 2023)
Vehicular/ITS	DRL signal+routing, adaptive	34% delay↓ (vs. static crossing+fixed route)	(Mushtaq et al., 2022)

These results demonstrate the universal applicability and the quantitative significance of dynamic traffic steering policies. When properly designed with real-time adaptation, RL or policy banks, and per-application awareness, such policies achieve near-optimal network performance across traffic scenarios, architectures, and application types.

---

References:

• "Dynamic Traffic Allocation in IEEE 802.11be Multi-link WLANs," (López-Raventós et al., 2022)
• "Packet Steering Mechanisms for MLO in Wi-Fi 7," (Cena et al., 2024)
• "Policy Reuse for Communication Load Balancing in Unseen Traffic Scenarios," (Xu et al., 2023)
• "Traffic Steering for 5G Multi-RAT Deployments using Deep Reinforcement Learning," (Habib et al., 2023)
• "Cascade Reinforcement Learning with State Space Factorization for O-RAN-based Traffic Steering," (Sun et al., 2023)
• "Machine Learning-enabled Traffic Steering in O-RAN: A Case Study on Hierarchical Learning Approach," (Habib et al., 2024)
• "A Deep Reinforcement Learning Approach for Adaptive Traffic Routing in Next-gen Networks," (Abrol et al., 2024)
• "On Deep Reinforcement Learning for Traffic Steering Intelligent ORAN," (Kavehmadavani et al., 2023)
• "Traffic Management of Autonomous Vehicles using Policy Based Deep Reinforcement Learning and Intelligent Routing," (Mushtaq et al., 2022)
• "Global QoS Policy Optimization in SD-WAN," (Quang et al., 2023)
• "Energy-Sustainable Traffic Steering for 5G Mobile Networks," (Zhang et al., 2017)
• "Intelligent Traffic Steering in Beyond 5G Open RAN based on LSTM Traffic Prediction," (Kavehmadavani et al., 2022)
• "On-Device Intelligence for 5G RAN: Knowledge Transfer and Federated Learning enabled UE-Centric Traffic Steering," (Zhang et al., 2023)
• "Optimal Routing for Delay-Sensitive Traffic in Overlay Networks," (Singh et al., 2017)