Improving Stability and Scalability in Edge Computing Task Scheduling Using an Adaptive Multi-Objective Evolutionary Algorithm

The rapid spread of latency-critical and computationally heavy applications in IoT, 5G, or the upcoming 6G has turned edge computing into more than just a buzzword it is now a foundational piece of the infrastructure. That said, the very features that make edge appealing (its distributed nature, dynamic behavior, and tight resource limits) also create serious headaches when it comes to scheduling tasks. Two pain points stand out: keeping performance stable when workloads fluctuate, and making sure the system can scale up without breaking as the network grows. Conventional scheduling approaches, along with static multi‑objective optimizers, tend to buckle under these pressures. Their performance often wavers unpredictably, and when you push them to larger setups, the computational cost skyrockets. To get a handle on this, we have designed an Adaptive Multi‑Objective Evolutionary Algorithm (AMOEA) that builds on a beefed‑up version of NSGA‑II. The core idea is to make the algorithm smart enough to adjust itself on the fly. Specifically, we have added three mechanisms: first, a dynamic tuning scheme for mutation and crossover rates so that exploration and exploitation stay balanced second, a continuous monitoring loop that watches workload changes and triggers adaptive relearning whenever needed, which helps maintain stability even when conditions shift and third, a diversity preservation trick that stops the population from converging too early, which in turn improves scalability. We ran a wide range of experiments, covering different workload patterns and system sizes, to see how well this approach holds up. The numbers speak for themselves: compared to classic NSGA‑II, MOPSO, and common heuristic methods, our AMOEA consistently comes out ahead. Performance variance dropped by as much as 46%, which is a strong sign of improved stability. More importantly, as the number of tasks increases, the latency growth curve is noticeably flatter so scaling really does work better. On top of that, we recorded meaningful gains across other metrics too: energy consumption fell by 8.15%, average response time shortened by 13.1%, and CPU utilization settled at a healthy 0.86 rate. The Hypervolume indicator, at 0.84, further confirms that the trade‑offs among objectives are well balanced. All in all, these results suggest that adding adaptive intelligence to evolutionary scheduling is a promising path forward especially for heterogeneous edge environments where conditions rarely stay the same for long. The framework we have put together appears robust, scalable, and efficient enough to handle the kind of dynamic resource management that real‑world deployments actually demand.