Introduction
As wireless communication systems transition toward 6G, the complexity of network management and optimization has increased exponentially. To facilitate the design and testing of AI-driven solutions without risking the stability of live production environments, researchers have turned to Digital Twins (DTs). A Digital Twin serves as a high-fidelity virtual replica of the physical network, enabling safe experimentation and performance prediction.
However, the utility of a Digital Twin is fundamentally limited by its accuracy in replicating real-world behavior. This process, known as "field data replication," involves calibrating simulator parameters—such as traffic patterns, user distributions, and resource allocation settings—so that the simulator's output aligns with real-world Key Performance Indicators (KPIs). This paper presents a novel approach developed at the Samsung AI Center that utilizes Mixed-Variable Particle Swarm Optimization (PSO) combined with $\alpha$-fairness to achieve high-accuracy, multi-objective replication.
The Replication Challenge
Field data replication in wireless networks faces two primary technical hurdles:
- Mixed-Variable Search Spaces: Network simulators are governed by a combination of continuous variables (e.g., packet sizes, power levels) and discrete variables (e.g., the number of users per cell, frequency band selections). Traditional optimization algorithms often struggle to navigate these hybrid spaces efficiently.
- Conflicting Objectives: Replicating a network requires matching multiple KPIs simultaneously, such as Cell Load and Downlink Volume. In many scenarios, these objectives are in conflict; optimizing for one may lead to significant errors in another. Standard weighted-sum approaches often require tedious manual tuning and can lead to "objective starvation," where certain metrics are neglected to minimize the total error.
Proposed Methodology
To address these challenges, the authors propose a framework that treats the network simulator as a black box, ensuring the method is applicable across various simulation platforms without requiring internal code modifications.
Mixed-Variable Particle Swarm Optimization
While standard PSO is effective for continuous optimization, it lacks a native mechanism for discrete jumps. The researchers implemented a specialized encoding scheme that represents continuous and discrete variables as distinct segments within a single "particle" position. This allows the swarm to explore the multi-dimensional parameter space holistically, evolving both types of variables in parallel to find the optimal configuration for the Digital Twin.
$\alpha$-Fairness Integration
To resolve the conflict between multiple KPIs, the authors integrated the concept of $\alpha$-fairness into the optimization objective. Traditionally used for resource allocation among users, $\alpha$-fairness is applied here to the KPIs themselves. By setting $\alpha=1$ (proportional fairness), the objective function ensures that the optimization effort is distributed equitably across all metrics. This prevents the algorithm from achieving a near-perfect match for one KPI at the expense of others, instead finding a balanced "sweet spot" that minimizes replication error across the entire set.
Experimental Setup and Results
The methodology was validated using real-world field datasets collected over a 24-hour period from five distinct geographic sites. The data was processed in 15-minute intervals, focusing on three critical KPIs: Active User Equipments (UEs), Cell Load (PRB utilization), and Downlink Volume.
The proposed Mixed-Variable PSO with Fairness was benchmarked against Random Search, Bayesian Optimization (BO), and Standard PSO. The results demonstrated significant improvements across three key dimensions:
- Accuracy: The proposed method achieved a Mean Absolute Percentage Error (MAPE) of 20.39%, outperforming both BO (20.76%) and Standard PSO (22.54%).
- Fairness and Balance: Using Jain’s Index—a metric where 1.0 represents perfect equality—the proposed method scored 0.80, indicating a much more balanced replication of KPIs compared to BO (0.74) and Standard PSO (0.75).
- Convergence Speed: Efficiency is critical when dealing with computationally expensive network simulations. The proposed algorithm reached a stable optimal state in just 7 iterations, whereas standard PSO required 8, and other methods exhibited lower stability over 50 iterations.
Discussion and Future Work
The integration of $\alpha$-fairness effectively eliminates the "manual weight trap," providing a mathematically rigorous way to balance conflicting objectives without human guesswork. Furthermore, the use of PSO particles allows for parallel exploration of the search space, leading to faster convergence than sequential methods like Bayesian Optimization.
Despite these successes, the authors identify several avenues for future research. Current optimization is performed on sector-level averages; moving toward UE-level replication would provide a higher-fidelity Digital Twin, albeit at the cost of significantly increased complexity. Additionally, the researchers aim to explore more advanced search mechanisms to enhance global exploration and prevent the algorithm from settling in local optima. Finally, scaling the framework to handle dozens of simultaneous KPIs will be essential for the next generation of ultra-dense 6G networks.
Conclusion
By combining mixed-variable optimization with a fairness-driven objective function, this research provides a robust framework for creating accurate Digital Twins. This approach ensures that virtual replicas are not only precise but also balanced across all critical performance metrics, paving the way for more reliable AI testing and network optimization in the 5G and 6G eras.
Summary updated