My Adventure
Fabio Berberi — Computer Science, Algorithms & Artificial Intelligence
I design and develop algorithmic solutions to complex problems, combining mathematical modeling, optimization techniques, and artificial intelligence.
My academic journey has been shaped by a strong interest in optimization, artificial intelligence, and complex systems. I began my studies in Italy at the University of Siena, where I developed a solid foundation in mathematics, engineering, and computational methods. During my Erasmus experience in Germany at Leuphana University of Lüneburg, I expanded my perspective by working in an international environment, collaborating with students and researchers from different backgrounds, and deepening my understanding of control systems and advanced modeling. Along this path, I have independently developed multiple research-oriented projects, focusing on optimization algorithms, Particle Swarm Optimization (PSO), and machine learning applications. My work has led to contributions in different domains, ranging from neural network optimization to pandemic modeling and safety-critical systems. In parallel, I have been involved in industrial-oriented problem solving, including work related to data analysis and error pattern detection in production systems, aligning with real-world challenges faced by companies such as Volkswagen and advanced manufacturing environments. My goal is to bridge theoretical research with practical impact, developing scalable and efficient algorithms capable of addressing complex, high-dimensional problems in both academic and industrial contexts.
Publications
[1] Domain-as-Particle with PSO Methods for Neural-Network Feature
Abstract
We present a framework that integrates Particle Swarm Optimization (PSO), machine learning, K-Fold cross-validation, and surrogate modeling to identify optimal weight vectors for feature scaling in neural network training. In our approach, the n-dimensional weight space is partitioned into non-overlapping subdomains, each corresponding to a PSO particle. Particle movement is guided by a characteristic vector determined by the best-performing candidates in each subdomain and by information exchanged with neighboring regions. To reduce evaluation costs, a surrogate model—trained on a uniformly sampled subset of candidates—pre-filters particles before full K-Fold validation. The top candidates then undergo comprehensive validation, updating the characteristic vectors for subsequent iterations. This domain-as-particle PSO framework enables efficient weight discovery, significantly reducing computational overhead while maintaining robust performance.
DOI: http://dx.doi.org/10.15439/2025F1427
FedCSIS — Krakow, Poland
Download PaperDescriptive Figures
[2] Optimal Vaccination Strategies for Pandemic Control: A Cost-Driven SIR Model with Domain-as-Particle PSO Optimization
Abstract
This chapter extends and improves upon a previously published SIR-based pandemic model with feedback vaccination law, which established a sufficient condition for achieving herd immunity through the minimization of a cost function combining both vaccination effort and intervention time. In the original approach, optimization was performed using standard routines, which proved to be computationally demanding and potentially limited in high-dimensional or nonlinear scenarios. Here, we introduce an advanced Particle Swarm Optimization (PSO) strategy based on a domain-as-particle paradigm, in which the parameter space is partitioned into non-overlapping subdomains, each acting as an independent PSO particle. This approach enables structured exploration of the solution space and enhances both convergence speed and robustness. The hybrid framework, integrating Simulink-based dynamic simulation with the domain-as-particle PSO, is demonstrated to achieve herd immunity more rapidly and efficiently compared to the original method, offering improved guidance for public health policy design.
Conference: icSoftComp2025 — Hanoi, Vietnam
Proceedings: Springer Proceedings
Download PaperDescriptive Figures
[3] Crash Scenario Optimization using Domain-as-Particle PSO
Abstract
The safety assessment of vehicles consists of both active and passive systems and crash conditions. Physical crash testing provides reliable information, yet it is costly, and large-scale numerical simulations quickly become computationally expensive. Therefore, key crash parameters such as impact velocity, collision angle, vehicle mass ratio, and structural stiffness are modeled as optimization variables, building upon previous work in multi-objective PSO-based crashworthiness optimization. Optimization is driven toward identifying severe crash scenarios through injury-related metrics, including the Head Injury Criterion (HIC) and chest deformation. Particle Swarm Optimization (PSO), originally introduced in the literature, has been widely applied in engineering optimization, including crashworthiness design and safety evaluation. In this work, we propose a Domain-as-Particle PSO (DaP-PSO) approach, where the parameter space is partitioned into non-overlapping subdomains, each acting as an independent particle. This structured formulation improves search efficiency, robustness, and convergence compared to classical PSO. Furthermore, recent advances in surrogate-assisted and data-efficient optimization highlight the importance of reducing expensive evaluations in simulation-driven problems. The simulation results demonstrate that DaP-PSO requires significantly fewer evaluations for detecting critical crash scenarios compared to classical PSO, making it a computationally efficient solution for virtual vehicle safety assessment.
Conference: ICCI 2025 — India
Conference: ICCI 2025 — Surat, India (presented in Surat, India)
Download PaperDescriptive Figures
[4] TGWO-SA: A Territorial Grey Wolf Optimizer Enhanced by Simulated Annealing, Perturbation Dynamics, and Natural Selection
Abstract
In this work, we introduce a refined version of the Grey Wolf Optimizer algorithm called TGWO-SA. The method integrates Simulated Annealing as a refinement phase, enabling improved local search capabilities. The algorithm incorporates adaptive population sizing and a communication mechanism inspired by wolf-pack cooperation, together with periodic elimination of poorly performing individuals, introducing a natural selection dynamic. Additionally, stochastic perturbations are applied to the best solution during local exploration, enhancing diversification and avoiding premature convergence. The algorithm is evaluated on benchmark functions and compared with classical optimization techniques such as GWO, PSO, and Differential Evolution (DE). Results show strong convergence stability and improved exploration performance. The goal is to develop a robust hybrid optimization framework capable of efficiently handling complex, high-dimensional search spaces.
Keywords: Grey Wolf Optimizer · Simulated Annealing · Hybrid Optimization · Perturbation · Natural Selection
Conference: ICCI 2025 — Surat, India (presented in Surat, India)
Download PaperDescriptive Figures
[5] Event-Triggered Robust Kalman Filtering with Firefly Optimization, Conformal Triggering, and Sparse Sensor Gating for Healthcare Monitoring
Fabio Berberi — University of Siena, Italy
Paolo Mercorelli — Leuphana University of Lüneburg, Germany
Abstract
Efficient and reliable monitoring of healthcare signals is a key challenge for edge computing and IoT-based clinical systems. We propose a novel event-triggered Kalman filtering framework enhanced by multiple advanced components. The method integrates Firefly-based initialization and bilevel energy–accuracy optimization, robust innovations with conformal triggering and change-point adaptation, and adaptive noise tuning via neural modulators combined with online EM and stability projection. Additionally, the framework incorporates sparse sensor gating using multi-armed bandits and Gumbel-Softmax selection, along with efficient computation through low-rank Riccati updates and mixed-precision quantization. A Koopman-lite encoding is used to approximate nonlinear dynamics, while federated and privacy-preserving deployment ensures applicability in real-world healthcare environments. Experimental results on real-world datasets (MIT-BIH ECG, PhysioNet PPG) and synthetic healthcare data demonstrate that the proposed method reduces update rates by up to 70%, achieves latency below 3 ms on CPU devices, and maintains robustness under noise, missing data, and sensor faults.
Keywords: Kalman Filter · Firefly Algorithm · Event-triggered filtering · Healthcare Monitoring · Sensor Fusion · Conformal Prediction
Download PaperDescriptive Figures
