The largest reported gains for a proposed wireless-networking method came in simulations of vehicular and adversarial networks: packet delivery ratio (PDR)—the share of message generations decoded successfully—was reported as 5.2 percentage points higher in the vehicular comparison and 9.8 points higher when malicious nodes were present, both against the cited comparison baselines. The paper also reports lower latency and higher node retention, but its evidence comes from modeled networks and a 10-node hardware testbed.
The system, called APC-RLNC, forms reliability clusters, applies tailored intra-cluster RLNC redundancy and uses hierarchical inter-cluster coding for cluster-level protection. In plain terms, it is a decentralized way of placing extra recovery information according to local link conditions, with a second layer to protect the clusters themselves.
What the model and simulations show
The paper’s worked decoding example reports approximately 97.1% success for any single cluster and approximately 0.99999998 system-level reliability when five independent clusters are combined. That is an illustrative calculation under specified channel parameters and independent-cluster assumptions, not a general guarantee for wireless networks.
The theoretical analysis also reports O(√T) regret for the online clustering rule. Here, regret is the gap between an online strategy and a benchmark over time; the paper states that the bound grows with the square root of the number of rounds under its assumptions. An illustrative estimate gives a cost of about 5,000 operations per epoch for 100 nodes, a generation size of 32 and about 10 clusters. That cost is an example, not a measured result for every configuration.
The evaluation compared APC-RLNC with Random, Static Clustering, PACE and ARLNC, using PDR, latency, overhead and retention. Each simulation ran for 500 steps, and results were averaged over 50 independent runs per configuration.
Under nominal conditions, APC-RLNC was reported at 97.82 ± 0.5% PDR, 142 ± 9 milliseconds latency, 18.6% overhead and 78.3% retention. Compared with Random, the text reports a 0.8-percentage-point PDR gain, a 0.82% relative gain, 10% lower latency and 20% higher retention. A t-test was reported as p < 0.01; confidence intervals and any adjustment for multiple comparisons were not reported in the supplied analysis.
In the vehicular scenario, APC-RLNC reached a reported 99.32 ± 0.6% PDR, 145 ± 11 ms latency and 82.1% retention. Its reported advantage over Random was 5.2 percentage points in PDR, or 5.5% on a relative basis, alongside an 18% latency reduction. The table gives standard deviations, but no scenario-specific p-value or confidence interval.
Not every test separated the methods
Bursty errors produced a more complicated picture. APC-RLNC maintained near-perfect delivery across the evaluated burst durations, but Static showed similar ceiling performance. Random fell to approximately 60.6% PDR when the burst lasted 50 packets. Because two approaches were already near the ceiling, this result does not by itself establish how much of the performance came from adaptive clustering.
In an adversarial simulation with 30 nodes and 20% malicious nodes, APC-RLNC with isolation reported 98.02 ± 0.7% PDR and 151 ± 12 ms latency. Random without defense reported 88.24 ± 2.1% PDR and 221 ± 24 ms, and the paper describes the difference as a 9.8-percentage-point gain over non-clustered baselines.
The largest reported scalability test used 500 nodes. Clustering took 42 milliseconds per epoch, described as 3% of a 1.5-second epoch, while reconfiguration overhead remained below 3% of total runtime. The analysis describes the distributed algorithm’s scaling as linear-in-N log N.
The energy calculation also favored APC-RLNC in the reported comparison: 4.21 joules per successfully delivered generation, versus 4.82 J for Random and 4.61 J for Static, a stated 12.7% reduction against Random. No variability estimate or inferential test for energy was reported.
A small hardware check
On hardware, the study used 10 Jetson Nano nodes with 2% emulated packet loss. It reported 96.91 ± 1.2% PDR, 156 ± 18 ms latency, throughput of 1.18 ± 0.15 Mbps and CPU use of 7.8 ± 1.1%. Reported errors against simulation were 0.93% for PDR, 9.86% for latency, 4.84% for throughput and 8.33% for CPU, all below 10%.
That check was reported to closely match the simulation on those measures, but it does not settle how the system would behave at larger scale. The 500-node result was simulated, while the hardware test involved 10 nodes; the number of independent testbed repetitions was not reported.
The boundaries of the evidence
The authors identify fairness, security, rapidly changing channels and redundancy scaling as limitations. The fairness objective uses average PDR, so persistently unreliable nodes may receive too little protection. The stated Byzantine-resilience mechanism does not provide stronger cryptographic guarantees, and extreme Doppler conditions may violate the convergence assumptions of the link-reliability tracker.
The paper also warns that its redundancy rule can over-provision stable clusters and create a ceiling effect in burst-error tests, where Static redundancy performs similarly. These caveats make the reported gains conditional on the channel and network conditions represented in the evaluation.
The document is a preprint labeled arXiv:2608.26040v1 and dated 26 August 2026. It states that the code, simulation configurations, evaluation scripts, testbed orchestration manifests and datasets used for the reported figures and tables are publicly available.
Paper data and sources
Original title: Adaptive Peer Clustering with Hierarchical Random Linear Network Coding for Resilient Decentralized Wireless Networks
Authors: Navaneetha Krishnan Kamalakannan, Harinisri Velmurugan
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text