Multi-Objective Edge Resource Optimization with D3QN-PER in Quantum-Classical IoT Edge Networks

Intelligent resource management is necessary for modern quantum-classical edge networks to manage diverse computational workloads while upholding stringent latency guarantees. We introduce D3QN-PER, a Dueling Double Deep Q- Network with Prioritized Experience Replay. It is intended for adaptive resource allocation on the edges of Quantum Key Distribution (QKD) post-processing pipelines with dynamic task distributions. With three diverse nodes handling bimodal workload distributions (88% baseline tasks, 12% computational spikes), our OpenAI Gym-based simulation environment simulates realistic edge scenarios. Through methodical two-stage training (2000 episodes, 300,000 scheduling decisions), D3QN-PER demonstrates exceptional performance: prevents catastrophic latency spikes (9× reduction in overload scenarios) during training, 40% overload mitigation, and total catastrophic failure elimination. We also performed rigorous 5-seed validation (1.5 million total decisions) to ensure reproducibility and confirmed robust performance: D3QN-PER (−73.61 ± 0.60) significantly outperforms Round-Robin (–152.07 ± 5.60, 51.60% improvement, d = 19.70, p < 0.001) and Least-Loaded (−187.97 ± 3.05, 60.84% improvement, Cohen’s d = 52.11, p < 0.001). The Dueling architecture contribution is isolated when compared to the D3QN-Uniform baseline (d = 0.17), and deployment profiling shows a GPU inference latency of 0.33 ms that is appropriate for edge hardware. In addition to offering the research community validated architectures, thorough hyperparameter configurations, and open simulation tools for future hardware deployment, this work establishes a reproducible framework for RL-driven edge scheduling with applications to IoT edge systems including quantum-classical networks, industrial automation, and smart cities.