HyperFusionNet: An Intelligent DRL Framework with Neural Attention and Memory Integration for Optimized IoT Communication and Energy Efficiency

Abstract

The sheer increase in IoT networks has created major challenges in the effective management of data communication, and high latency, limited power supply resources, and unreliable networks are some of the problems that have been encountered. All these limitations pose challenges to the scalability and performance of the IoT system, and more sophisticated data routing, energy consumption, and network stability optimization solutions are required. In this study, it is suggested to use HyperFusionNet, a progressive Deep Reinforcement Learning (DRL) model aimed at maximizing communication in a large-scale IoT network. HyperFusionNet extends the use of the attention and memory modules in neurons with the objective of reducing latency, improving throughput and increasing network lifetime, i.e., tackling the limitations of energy and large loads of data in IoTs. The framework was evaluated over 100 simulated episodes and converged within 80 episodes, achieving a normalized average cumulative reward of 960. Across the evaluated conditions, HyperFusionNet maintained latency of 90-120 ms, a throughput of 900-1200 bps, PDR of 95-99%, and an average network power of 1000-1050 mW. Under the common experimental pipeline, it provided approximately 15% lower latency, 12% higher throughput, and a 20% improvement in network lifetime relative to the internal reference configuration. Also, the energy used per node was reduced by almost one-fifth that of other routing protocols, increasing life span of the network by nearly one-fifth. The DRL agent used in HyperFusionNet is the Proximal Policy Optimization (PPO) that has curiosity-based exploration and reward shaping strategy to adapt to real-time network conditions. The model converged in 80 episodes with an average cumulative reward of 960 and a robustness of their policy that retained 95% reliability in the unfavorable conditions, e.g., higher traffic and node failures.