Adaptive Peer Clustering with Hierarchical Random Linear Network Coding for Resilient Decentralized Wireless Networks

Navaneetha Krishnan Kamalakannan; Harinisri Velmurugan

Source-linked AI summary

Adaptive Peer Clustering with Hierarchical Random Linear Network Coding for Resilient Decentralized Wireless Networks

Navaneetha Krishnan Kamalakannan, Harinisri Velmurugan

arXiv:2608.26040v1Aug 26, 2026eess.SPcs.DCcs.ITcs.NI

TL;DR

Heterogeneous, dynamic wireless networks need robust communication despite mobility, intermittency, interference, energy limits, and scale. APC-RLNC clusters peers by reliability and applies hierarchical RLNC, with evaluations reporting PDR and latency gains across challenging scenarios, linear complexity to 500 nodes, and low reconfiguration overhead.

Problem
Dynamic topologies, heterogeneous link quality, transient failures, interference, power limits, and large peer populations challenge robust communication in decentralized wireless collectives.
Method
APC-RLNC dynamically clusters peers by link reliability and exchanges RLNC-encoded summaries across clusters for secondary protection.
Results
Evaluation reports 5.2% PDR gain and 18% latency reduction in high-mobility vehicular networks, 9.8% PDR gain under adversarial interference, and under 3% reconfiguration overhead to 500 nodes.
Takeaways & Limitations
The analysis and evaluation support adaptive clustering as a way to improve RLNC performance in heterogeneous wireless environments.
Takeaways & Limitations
The redundancy policy can over-provision under relatively stable cluster sizes and evaluated burst-error settings, causing near-perfect decoding that limits visible adaptive gains.

Abstract

from arXiv · show

Decentralized wireless collectives including vehicular swarms, IoT clusters, and edge AI networks require communication protocols that maintain robustness under dynamic topologies and heterogeneous link quality. While Random Linear Network Coding (RLNC) provides algebraic resilience against packet erasures, its performance degrades significantly when peers exhibit diverse channel conditions. This paper presents Adaptive Peer Clustering with Hierarchical RLNC (APC-RLNC), a system that dynamically groups peers by exponentially weighted moving average (EWMA) reliability metrics and applies multi-tier network coding within and across clusters. We formalize the clustering optimization problem, derive closed-form decoding probability bounds for Markov erasure channels, and prove O(sqrt(T)) regret for online reconfiguration under the Follow-the-Regularized-Leader (FTRL) framework. Our implementation includes both a high-fidelity network simulator and a proof-of-concept testbed deployment on Jetson Nano edge devices. Evaluation across diverse scenarios including high-mobility vehicular networks, burst-error channels, and adversarial interference demonstrates 5.2-9.8 percentage-point packet delivery ratio (PDR) improvements, 10-23% latency reductions, and up to 30% higher node retention compared to state-of-the-art baselines. The system exhibits linear scalability to 500+ nodes and maintains real-time reconfiguration overhead below 3%. APC-RLNC establishes adaptive clustering as a foundational primitive for AI-native 6G wireless systems.

I. INTRODUCTION

Decentralized wireless networks must remain robust despite heterogeneous links, dynamic topology, energy constraints, and interference. APC-RLNC addresses these challenges by clustering peers by reliability and combining adaptive local redundancy with hierarchical protection.

Motivation: Heterogeneous erasure rates, frequent churn, and sub-5 W energy constraints complicate decentralized coordination across hundreds to thousands of peers.Typical churn reaches 2–10% per transmission round.
Motivation: Classical RLNC assumes homogeneous erasure rates, causing redundancy waste for reliable peers and insufficient protection for unreliable peers.These mismatches can produce throughput collapse and fairness violations.
APC-RLNC: APC-RLNC clusters peers by link reliability, then applies tailored intra-cluster RLNC and secondary inter-cluster coding.The design uses EWMA tracking, online clustering optimization, and hierarchical encoding and decoding.
APC-RLNC: High-reliability clusters receive minimal coding overhead while low-reliability clusters receive proportional protection.This targets localized redundancy optimization without global coordination overhead.
APC-RLNC: Distributed clustering has O(N log N) complexity, and intra-cluster RLNC operations parallelize trivially.The framework is implemented in a simulator and on Jetson Nano devices.
Evidence: The paper derives Markov-channel decoding bounds, proves O(sqrt(T)) FTRL regret, and evaluates the system against uniform-redundancy and centralized-feedback baselines.Evaluation covers nominal, high-mobility, burst-error, and adversarial scenarios.

B. Peer Clustering in Distributed Systems

Prior clustering and coding approaches organize peers by proximity, content, topology, or centralized allocation rather than continuously adapting redundancy to channel quality. This paper frames decentralized networks with time-varying links, churn, and vehicular mobility as the operating setting.

Prior Approaches: Earlier P2P clustering schemes grouped peers by proximity or content affinity, not channel quality.APC-RLNC explicitly clusters by erasure statistics for redundancy optimization.
Prior Approaches: Opportunistic routing and rate adaptation adjust per-link parameters but retain fixed topology granularity, while hierarchical coding uses static assignments.APC-RLNC continuously adapts clustering to time-varying channels with regret bounds.
System Context: AI-native wireless systems emphasize decentralized intelligence, ultra-low latency, and resilience, with APC-RLNC positioned as a coding layer beneath higher-level ML policy.The framework can deploy as an O-RAN RIC xApp.
System Model: The network model uses an undirected time-varying graph with link-specific erasure probabilities, node churn, and a Poisson per-step churn rate of 0.02–0.10.Nodes depart and join to maintain an expected network size.
System Model: Vehicular scenarios use random-waypoint motion with velocities uniformly distributed from 5 to 30 m/s.The model couples mobility with distance-dependent channel behavior.
Coding Background: RLNC encodes K original packets into K + R coded packets, and decoding requires at least K linearly independent packets.This coding primitive underlies the paper’s adaptive redundancy design.

C. Heterogeneous Channel Models

APC-RLNC models channel heterogeneity through Bernoulli, Markov, and velocity-dependent fading processes, then uses EWMA reliability scores to form clusters and allocate local redundancy.

Channel Models: The evaluation considers Bernoulli, two-state Markov, and velocity-dependent fading channel models.These models cover independent variation, burst-like state changes, and mobility-linked fading.
Channel Models: Bernoulli channels assign each node an i.i.d. erasure probability uniformly distributed from 0.05 to 0.25.The model captures heterogeneous but independently varying node reliability.
Channel Models: Markov links alternate between good states with p_g = 0.05 and bad states with p_b = 0.40, using transition probabilities P_gb = P_bg = 0.10.The state transitions model temporally correlated errors.
Channel Models: Velocity-dependent fading sets p_i(t) = p_base + β · |v_i(t)| with β = 0.003 s/m.The term models Doppler effects in vehicular networks.
Reliability and Clustering: Each node maintains an EWMA reliability score from instantaneous packet success rate, with smoothing factor α = 0.2.The score estimates 1 − p_i(t).
Reliability and Clustering: Peers partition into clusters by penalizing intra-cluster heterogeneity while rewarding larger clusters.This balances reliability homogeneity against over-fragmentation.
Hierarchical RLNC: Within each cluster, local RLNC uses redundancy based on cluster size and average erasure probability, while bridge packets provide cross-cluster protection.A global decoder can recover a generation when at least K total degrees of freedom arrive across clusters.

IV. THEORETICAL ANALYSIS

The theoretical analysis characterizes hierarchical decoding under independent clusters and illustrates how bridge coding amplifies system reliability. A worked Markov-channel example also validates the corrected per-cluster probability against Monte Carlo simulation.

Hierarchical Decoding Probability: For independent clusters, global decoding fails only when all clusters fail, yielding a system probability obtained from the product of cluster failure probabilities.The theorem provides the basis for hierarchical decoding reliability.
Worked Example: In the worked Markov example, the steady-state average erasure rate is 0.225 for K = 32 and R_c = 16.The example uses equal good and bad state probabilities.
Worked Example: Approximately 97.1% per-cluster success combines across five clusters to produce P_sys ≈ 0.99999998.The example demonstrates the value of bridge coding despite moderate per-cluster reliability.
Validation and Correction: The corrected ≈0.971 per-cluster probability matches 4,000-trial Monte Carlo simulation within 0.2%.An earlier draft’s ≈0.989 value corresponded to different redundancy parameters.

B. Online Optimization Regret

The online clustering policy models assignments on a simplex and uses FTRL under convex, Lipschitz losses to control cumulative regret. The section also reports computational costs for clustering, encoding, and decoding.

FTRL operates on clustering assignments xt ∈∆C, with loss ℓt(xt) measuring clustering quality.
O(sqrt(T)) regret is proved for online clustering reconfiguration under convex loss relaxation and Lipschitz constant G.
With G = 0.05 and T = 500 steps, the bound is negligible relative to cumulative loss and matches the empirical regret trace.
O(Nk) clustering, O(CK2) encoding, and O(K3) per-cluster decoding characterize the principal computational costs.
For N = 100, K = 32, and C ≈10, per-epoch cost is approximately 5,000 operations, below real-time constraints on modern edge devices.

V. APC-RLNC SYSTEM DESIGN

APC-RLNC uses lightweight reliability state and periodic distributed clustering to organize peers into adaptive groups. Gossip-based consensus updates local cluster estimates and assignments over repeated rounds.

System architecture: Nodes estimate reliability from per-packet acknowledgments, smooth measurements with Eq. (4), and periodically broadcast 2-byte state.
Distributed clustering: Every τ steps, gossip-based consensus estimates per-cluster means and variances and converges in O(log N) rounds.
System architecture: The architecture tracks link reliability with EWMA, clusters peers every τ epochs, and supports hierarchical encoding and decoding within and across clusters.
Distributed clustering: Algorithm 1 initializes cluster centers, exchanges reliability scores and cluster IDs, updates centroids, reassigns nodes, and merges or splits clusters when beneficial.

C. Hierarchical Coding Module

The coding module combines local RLNC generations with bridge packets that summarize clusters for cross-cluster recovery. It also incorporates energy-aware operation and defenses against malicious reliability reports.

Hierarchical coding: Cluster nodes encode local generations with Rc redundancy, while rotating representatives form bridge packets from cluster-generation summaries.
Hierarchical coding: Bridge redundancy scales with cross-cluster heterogeneity, and global decoding succeeds when at least K total degrees of freedom arrive across clusters.
Energy management: Nodes enter low-power mode when Ei < Ethreshold, reducing transmission duty cycle by 50%.
Hierarchical coding: Inner coding loops operate within clusters, while an outer loop generates bridge packets for cross-cluster resilience.
Adversarial resilience: Byzantine-resilient aggregation and anomaly-based cluster isolation address false reliability reports from malicious nodes.

VI. IMPLEMENTATION

The implementation combines a Python/NumPy discrete-event simulator with a Jetson Nano wireless testbed and records application, latency, and resource metrics. Simulation outputs agree closely with the analytical model, while experiments cover nominal and vehicular conditions.

Simulation platform: The simulator uses NetworkX for topology, NumPy for algebra, custom GF(256) arithmetic, and random-waypoint mobility with optional SUMO traces.
Simulation platform: A Cython extension provides a 5× speedup on the simulator’s encode/decode hot loop.
Simulation platform: 1% validation error bounds the simulator’s agreement with the analytical model of Eq. (3).
Evaluation setup: Each simulation run spans 500 steps, with 50 independent runs averaged per configuration.
Hardware testbed: The testbed uses 10 Jetson Nano devices, ad-hoc 802.11ac Wi-Fi, controlled packet loss, UDP gossip, mirrored GF(256)/RLNC libraries, and Kubernetes orchestration.
Evaluation setup: Latency, throughput, and CPU/memory usage are measured with kernel timestamps, application-layer logging, and perf profiling, respectively.

VII. EVALUATION

APC-RLNC is evaluated against fixed, adaptive, and centralized baselines across nominal, burst-error, and adversarial settings. It improves delivery, latency, retention, and burst resilience, while hardware comparison is included.

Experimental setup: The evaluation compares Random, Static Clustering, PACE, and ARLNC against APC-RLNC using PDR, latency, overhead, and retention.Random uses fixed R = 16; Static Clustering does not adapt after initialization, while PACE and ARLNC are reference baselines.
Nominal conditions: 0.8 percentage-point PDR gain, 10% latency reduction, and 20% higher retention are reported over Random under nominal conditions.The PDR gain is 0.82% relative, with statistically significant gains reported at p < 0.01.
High-mobility networks: 5.2 percentage-point PDR gain and 18% latency reduction are reported over Random in high-mobility conditions.Adaptive clustering updates every τ = 20 steps, approximately 12 seconds, to track mobility-induced erasure changes.
Burst-error channels: Approximately 60.6% PDR is reached by Random at burst duration 50, while APC-RLNC maintains near-perfect delivery under the evaluated Markov-channel conditions.Static redundancy shows similar ceiling performance because its fixed redundancy remains sufficient in the simulated loss conditions.
Hardware validation: Simulation and testbed measurements remain within 10% error, validating the reported metrics on a 10-node Jetson Nano deployment.The testbed comparison is identified as Table IV.
Adversarial interference: 9.8 percentage-point PDR gain over non-clustered baselines is reported under adversarial interference through Byzantine-resilient cluster isolation.The adversarial scenario contains 30 nodes and 20% malicious nodes.

F. Scalability (500 Nodes)

APC-RLNC scales to 500 nodes with low runtime overhead and lower energy use, but its largest gains depend on network heterogeneity, churn, and sufficient scale. Centralized methods retain optimality advantages while requiring global coordination.

Scalability: 42 ms per epoch at N = 500 represents 3% of a 1.5 s epoch, while the distributed algorithm follows linear-in-N log N complexity.The measured runtime is presented as evidence for real-time clustering at 500 nodes.
Energy: 12.7% lower energy per successfully delivered generation is reported for APC-RLNC than Random.Energy values are 4.21 J for APC-RLNC and 4.82 J for Random; reduced retransmissions lower total energy draw.
Operating regime: Clustering provides the largest gains when erasure heterogeneity is high, churn is moderate, and N ≳30 permits parallelism.The stated boundaries are variance in p_i above 0.05 and churn ρ_churn ∈ [0.02, 0.10].
Trade-offs: Centralized schemes achieve optimal allocations but require O(N^2) communication, single-point-of-failure exposure, and millisecond-scale control-plane latency.APC-RLNC trades approximately 1–2% versus an oracle for decentralization, fault tolerance, and sub-second adaptation.
Applications: APC-RLNC is positioned for federated learning, swarm robotics, and edge inference, with future work targeting joint clustering and ML-training optimization.The paper describes these as complementary application areas rather than evaluated deployment results.
O-RAN integration: In O-RAN, APC-RLNC can operate as a near-RT RIC xApp using link statistics and redundancy policies without physical-layer changes.OpenAirInterface prototyping reports sub-5 ms control-loop latency.

E. Limitations •

APC-RLNC improves RLNC performance in heterogeneous wireless environments, but its current design has fairness, security, channel-stationarity, and redundancy-scaling boundaries. The paper also reports feasibility at 500 nodes and on a 10-node Jetson Nano testbed.

Fairness: Average-PDR optimization may leave persistently unreliable nodes insufficiently protected.The authors identify max-min fairness constraints as future work.
Security: Byzantine-resilient aggregation tolerates fewer than 1/3 adversaries, while stronger cryptographic guarantees remain future work.
Non-stationary channels: Extreme Doppler shifts, including high-speed rail at 350 km/h, may violate EWMA convergence assumptions.Kalman-filter-based tracking is under investigation.
Redundancy scaling: Under stable cluster sizes and evaluated burst-error settings, Equation (6) can over-provision redundancy and create a near-perfect-decoding ceiling.Future work will examine normalized or budget-constrained redundancy allocation under moderate burst conditions.
Reported feasibility: APC-RLNC exhibits linear complexity to 500 nodes, under 3% reconfiguration overhead, and feasible Jetson Nano operation with at least 1 Mbps coded throughput.The 10-node hardware testbed measured under 8% CPU load and millisecond-scale reconfiguration latency.