Swarm-guided adaptive routing boosts energy efficiency in wireless IoT sensor networks

Hive Meets Deep Learning: Hybrid AI Routing Boosts IoT Sensor Networks’ Energy Efficiency by Up to 38%

The Internet of Things has a quiet addiction problem: it runs on batteries that nobody can easily replace. Billions of wireless sensor nodes — clipped to hospital patients, bolted to power-grid equipment, buried in farmland and forests — are expected to stream data around the clock, yet most will die not because their electronics fail but because their batteries empty. Now an international research team reports a candidate cure. In a study published on 29 August 2026 in the journal Cluster Computing, a group led by Mehdi Hosseinzadeh of Duy Tan University in Vietnam unveils ASGRR, short for Adaptive Swarm-Guided Graph Policy Routing, a framework that fuses graph neural networks, reinforcement learning and honeybee-inspired optimization into a single, self-tuning routing engine. In simulations, networks running ASGRR retained up to 38.06% more residual energy than those using competing routing schemes. The work brings together three strands of artificial intelligence that rarely share a single protocol stack, and the early numbers suggest the combination is more than the sum of its parts.

The problem ASGRR attacks is structural. Wireless sensor networks, the nervous system of the Internet of Things, juggle three compounding failure modes. Energy comes first: nodes are small, cheap and battery-powered, every transmission draws down reserves, and the weakest links deplete first — when a node dies, its neighbors inherit its routing burden and burn out faster, a cascade that drains whole regions of a deployment. Churn comes second: topologies shift constantly as nodes fail, move or sleep, severing routes that were optimal only moments before. Volatility comes third: traffic in healthcare monitoring, smart grids and environmental sensing surges and collapses unpredictably, so a policy calibrated for yesterday’s load misjudges today’s. Together, the authors note, these forces reduce reliability, raise latency and cap scalability — and they bite hardest in mission-critical scenarios that demand real-time delivery from heterogeneous devices whose behavior no one can fully predict.

The team behind the paper spans three continents: Hosseinzadeh, with dual affiliations at Duy Tan University in Da Nang, Vietnam and the Jadara University Research Center in Irbid, Jordan; Parisa Khoshvaght of the Saveetha Institute of Medical and Technical Sciences in Chennai and Western Caspian University in Baku; Amir Masoud Rahmani of National Yunlin University of Science and Technology in Taiwan; Mohammad Mohammadi of the ACECR higher-education institute in Khuzestan, Iran; Amin Mehranzadeh of Islamic Azad University’s Dezful branch; Thantrira Porntaveetus of Chulalongkorn University in Bangkok; and Sang-Woong Lee of Gachon University in South Korea. The manuscript was received on 10 August 2025, revised on 14 March 2026 and accepted on 4 July 2026, appearing as volume 29, article 709 of Cluster Computing. The authors report receiving no financial support for the research, authorship or publication of the study.

ASGRR’s first pillar is a Message Passing Neural Network, or MPNN — the graph-native cousin of the deep networks that transformed image recognition. Rather than consuming flat tables of numbers, an MPNN treats the network as what it truly is: a graph of nodes joined by links of varying quality. Each node carries a feature vector summarizing its state, such as residual energy and traffic load, and repeatedly exchanges these “messages” with its neighbors. After every exchange, each node aggregates the incoming information and updates its own representation, so after k rounds of message passing a node has effectively sensed conditions up to k hops away. ASGRR extends this machinery with a long-term memory module, a recurrent store that preserves the network’s topology history across time. By fusing present-tense messages with remembered history, the module captures complex spatial–temporal relationships between nodes and learns to predict topology changes — an impending link failure, a drifting node, a neighbor about to exhaust its battery — before they sever connectivity.

Prediction alone, however, routes nothing. The second pillar, Policy Gradient Reinforcement Learning, turns foresight into decisions. Policy gradient methods dispense with value lookup tables; instead, a neural network directly parameterizes the routing policy, actions are sampled from that policy, and the network’s weights are nudged uphill along the gradient of expected reward, with rewards shaped by residual energy, delay and delivery success. ASGRR sharpens this recipe with two mechanisms. A shared experience memory pools the trajectories recorded across the network, so each routing agent learns from the collective record of successes and failures rather than its own observations alone, accelerating learning and stabilizing training in sparse-reward conditions. Meanwhile, a predictive link stability metric, derived from the MPNN’s forecasts, feeds directly into the reward function, penalizing paths that thread through links the predictor expects to fail. The result is multi-criteria path selection that weighs energy, latency and link durability together, instead of the myopic shortest-hop logic that quietly bleeds sensor networks dry.

The third pillar sweats the local details. The Artificial Bee Colony algorithm is a swarm-intelligence method modeled on foraging honeybees, with three castes of artificial agents: employed bees that exploit known food sources, onlooker bees that probabilistically recruit toward the richest ones, and scout bees that abandon exhausted sources to probe new territory. ASGRR casts candidate paths as food sources and lets the colony locally refine the routes the policy-gradient layer proposes. Its signature twist is entropy-based dynamic clustering. Entropy — the statistical currency of disorder — is computed over local traffic and topology signals, and scout-bee allocation follows the result: where entropy spikes, a hallmark of churn, mobility or instability, more scouts are dispatched to explore alternative routes; where the neighborhood is calm, bees shift into exploitation, polishing the best-known paths with minimal overhead. The colony thus concentrates its search effort exactly where the network is least predictable.

What distinguishes ASGRR from earlier mash-ups of machine learning and nature-inspired optimization is that the hybrid is self-adaptive rather than hard-wired. The framework dynamically tunes its own parameters as conditions evolve, re-balancing the contributions of memory, policy and swarm search against real-time measurements of network volatility. In calm stretches, the system leans on accumulated experience and exploits proven routes; in turbulent ones, it widens exploration, extends its predictive reach and lets the bee colony roam. Architecturally, that means ASGRR behaves less like a static protocol and more like a closed control loop — continuously sensing, predicting and reorganizing. According to the authors, precisely this self-adaptation is what energy-constrained, dynamically changing WSN-IoT environments demand, where unpredictable node behavior, heterogeneous device capabilities and real-time delivery requirements conspire to defeat routing schemes that assume a stable world.

The benchmark numbers are the study’s loudest claim. In simulation analyses, ASGRR was pitted against three established baselines: MPSORP, BWOA+FIS — a clustering-and-routing scheme that couples a binary whale optimization algorithm with a fuzzy inference system — and FC-CRA. Against that field, ASGRR delivered up to a 38.06% improvement in residual energy, a margin that maps almost directly onto network lifetime, because a sensor’s death is first and foremost an energy event. Delay dropped by up to 14.88%, an effect consistent with the policy’s habit of steering traffic away from congested and unstable relays. And the packet delivery ratio — the fraction of generated packets that actually reach their destination, arguably the most user-visible metric in any sensing system — climbed by up to 13.49%. For a hospital telemetry network or a smart-grid control loop, a thirteen-point gain is not an academic curiosity; it is the difference between an alarm that arrives and one that silently never does.

The target applications are precisely the ones whose demanding traffic profiles motivated the work: healthcare, smart grids and environmental monitoring, domains where nodes are energy-starved, topologies move and data is mission-critical. The study also sits inside a recognizable research lineage from the same community — earlier work on fuzzy multi-criteria decision-making combined with enhanced random forests for body-area networks, fuzzy MAIRCA and eigenvector methods for IoT routing, and gray-wolf-plus-fuzzy clustering for WSN-IoT — a progression that reads as a deliberate campaign to replace hand-tuned heuristics with learning-based alternatives. The caveats, however, are the standard ones for the genre. The performance claims rest on simulation analysis, and the paper’s data availability statement notes that no datasets were generated or analysed during the study. Real deployments will bring radio irregularities, interference, security pressures and the tight memory and compute budgets of milliwatt-scale microcontrollers — and proving that a three-engine AI stack repays its own overhead in the field remains the decisive next test.

Even so, the trajectory is telling. For two decades, sensor-network protocols were designed by hand, tuned once and deployed blind; the emerging generation treats the network as a system that can observe its own history, anticipate its own failures and reorganize on the fly. ASGRR draws simultaneously on deep learning over graphs, policy gradient reinforcement learning and insect-inspired swarm optimization — three traditions that rarely meet in a single protocol — on the argument that no one tool can handle networks whose energy, topology and traffic all misbehave at once. The metaphor at the system’s heart may prove its most lasting contribution. A healthy beehive endures not because any individual bee is strong, but because the colony constantly re-allocates its scarce workers to wherever they matter most. If ASGRR’s simulated gains survive contact with real hardware, the Internet of Things may have acquired the same instinct: a routing layer that behaves less like fixed plumbing and more like a hive, quietly redistributing its scarce energy to keep the colony — and the data flowing through it — alive.