Alejandro Morales Hernández

Alejandro Morales Hernández

He/Him

Postdoctoral researcher

Talks and presentations

Ensemble Learning: Building Smarter Models through Collaboration

November 24, 2025

Talk, BRAIN - AI cafe / online workshop, Belgium

In supervised learning, a single model often struggles to capture all the underlying patterns of a complex problem. As tasks become more complex and the consequences of inaccurate predictions grow more significant, more sophisticated approaches are required. This challenge has motivated the development and success of ensemble systems, which combine multiple models to achieve higher accuracy, robustness, and generalization than any individual model. In this presentation, I will outline the fundamental components of ensemble learning and discuss the three principal architectures commonly found in the literature: Bagging, Boosting, and Stacking. Finally, three case studies will be presented (related to unbalanced learning, missing data imputation, and data fusion) to demonstrate how each of these architectures can be effectively applied to real-world problems in traffic and energy systems.

A Tutorial on Smart Process Optimization with Bayesian Optimization: Leveraging AI for Efficient Decision-Making

October 26, 2025

Tutorial, ECAI2025 - The Engineering School, University of Bologna, Bologna, Italy

In this tutorial, we introduce how Bayesian optimization (BO) can be applied to process-optimization tasks for researchers and practitioners, using an adhesive bonding process as example. We explain the theoretical foundations of BO and provide practical guidance (yes, with code) for implementing it in real-world workflows.

Leveraging SUMO for Traffic Twins: Experiences in Urban Traffic Processing

May 13, 2025

Poster, SUMO User conference - German Aerospace Center (DLR), Berlin-Adlershof, Germany

The TORRES project develops an AI-driven framework that integrates diverse mobility data sources with SUMO simulations to improve real-time traffic monitoring, prediction, and decision-making in urban environments. Applied in Brussels and Namur, the framework focuses on three key contributions: traffic calibration, data interpolation, and simulation-assisted prediction for efficient mobility management.

Integration of simulation and machine learning to predict the traffic in the event of disruptions

June 27, 2024

Poster, Rencontres Francophones Transport Mobilite (RFTM2024) - Université Libre de Bruxelles (ULB), Brussels, Belgium

Decision-making in intelligent transportation systems is complex due to urban traffic’s unpredictable dynamics, particularly when assessing the impact of road works on congestion. To address this, a method combining simulation using SUMO and machine learning with GNN is proposed for traffic prediction in the event of road disruptions.

Multi-objective optimization of adhesive bonding process in constrained and noisy settings

May 03, 2023

Talk, International Conference on Optimization and Learning (OLA2023), University of Málaga, Málaga, Spain

This research applies machine learning models within a Bayesian optimization framework to efficiently identify Pareto-optimal parameters for adhesive bonding processes, balancing strength, cost, and quality constraints. By emulating objective and constraint functions with limited experimental data, the approach reduces the need for costly and time-consuming lab experiments.

Common pitfalls in practical multi-objective optimization

March 30, 2023

Poster, Flanders Make: CMVPT Conference on Machines, Vehicles and Production Technology, Gent, Belgium

The literature on multi-objective optimization is vast. However, in practice not only the most commonly used approaches can yield bad quality solutions, but the noisy performance is often neglected and the problem is treated as deterministic. In this work we highlight some of the most crucial pitfalls when using common multi-objective algorithms in practice.

Encore Abstract: Online learning of windmill time series using Long Short-term Cognitive Networks

November 07, 2022

Talk, Joint International Scientific Conferences on AI and Machine Learning (BNAIC/BeNeLearn2022), Lamot Conference and Heritage Centre, Mechelen, Belgium

This work introduces Long Short-term Cognitive Networks (LSTCNs), a novel gated neural network designed for efficient online learning in windmill farm forecasting. Case study results show that LSTCNs achieve lower forecasting errors and faster training compared to traditional recurrent models.

Multi-objective hyperparameter optimization with performance uncertainty

July 18, 2022

Talk, International Conference on Optimization and Learning (OLA2022), Syracusa Faculty of Architecture, Syracusa, Sicilia, Italy

This paper proposes a multi-objective hyperparameter optimization method that combines Tree-structured Parzen Estimators with Gaussian Process Regression to account for uncertainty in performance evaluations. Experiments on analytical test functions and ML problems demonstrate improved optimization results compared to using TPE or GPR alone.

Multi-objective optimization of stochastic experiments

March 30, 2022

Poster, Flanders Make: CMVPT Conference on Machines, Vehicles and Production Technology, Gent, Belgium

Many engineering problems in areas such as system design, process optimization, and supply chains involve multiple competing objectives and noisy experimental outcomes. The relationships between design variables, objectives, and constraints are often black-box in nature, making evaluation expensive and forcing engineers to work with limited experimental budgets. To identify high-quality or near-optimal solutions efficiently, these challenges call for machine learning approaches tailored to scarce data, combined with optimization and statistical learning methods, rather than traditional heuristic algorithms.

Multi-objective simulation optimization of the adhesive bonding process of materials

December 13, 2021

Talk, Winter Simulation Conference (WSC2021), JW Marriott Phoenix Desert Ridge Resort & Spa, Phoenix, Arizona, US (virtual)

This research applies Bayesian optimization with Gaussian Process Regression and Logistic Regression to identify optimal adhesive bonding process parameters efficiently. The approach reduces experimental effort while guiding designs toward Pareto-optimal solutions for lightweight automotive applications.