Institution profile

Universidad Carlos III de Madrid

Academic institutioneurope · es
Official website
Research library253linked papers
Opportunities0open roles
Selected work

Representative Papers

An MDP Model for Censoring in Harvesting Sensors: Optimal and Approximated Solutions

Jan 14, 2015IEEE Journal on Selected Areas in Communications

This paper addresses energy-efficient information transmission for energy-harvesting sensors, aiming to maximize cumulative message utility (i.e., importance) under finite energy constraints. Method: We formulate the problem as an infinite-horizon Markov decision process (MDP) and—under a realistic battery dynamics model—rigorously prove that the optimal policy is a state-dependent importance-threshold truncation policy, where transmission decisions depend dynamically on the current battery level. Building on this structural insight, we propose a low-complexity, fast-converging model-driven stochastic approximation algorithm and benchmark it against Q-learning. Results: Experiments in both single-hop and multi-hop networks demonstrate that our algorithm significantly reduces computational overhead and accelerates convergence while achieving utility performance close to the theoretical optimum.

19 citations2 influentialRead paper

Fuzzy Model Identification and Self Learning with Smooth Compositions

Oct 03, 2019International Journal of Fuzzy Systems

To address slow convergence, severe oscillations, and learning instability in fuzzy system modeling—caused by rule discontinuities and parameter uncertainty—this paper proposes an end-to-end differentiable fuzzy modeling and self-learning framework based on smooth composition operators. The core innovation is the first introduction of a smooth T-norm and S-implication composite structure, enabling global differentiability of both membership functions and the inference process. This facilitates gradient-driven adaptive updating of fuzzy rules and joint optimization of antecedent and consequent parameters. Evaluated on multiple nonlinear system identification benchmarks, the method achieves substantial improvements: enhanced modeling accuracy and generalization capability, 40% faster convergence, and a 75% reduction in rule oscillation—without incurring significant computational overhead.

14 citationsRead paper

Machine Learning Inference on Inequality of Opportunity

Jun 10, 2022

This paper addresses the bias induced by machine learning (ML) plug-in estimators in U-statistics—particularly inequality of opportunity (IOp)—and proposes the first debiased IOp estimator with accompanying asymptotic inference theory. Methodologically, it integrates double robustness, U-statistic theory, and general-purpose ML prediction (e.g., tree-based models, regression). The approach effectively corrects estimation bias arising from model selection and regularization. Contributions include: (1) a novel debiasing framework for general U-statistics, overcoming the sensitivity of conventional plug-in methods to ML model specification; (2) the first unbiased cross-national IOp estimates for European countries; and (3) empirical identification of maternal education and paternal occupation as the most critical circumstances shaping IOp. Simulation studies demonstrate substantial improvements in estimation accuracy and reliable statistical inference.

7 citations1 influentialRead paper

SYMBXRL: Symbolic Explainable Deep Reinforcement Learning for Mobile Networks

May 19, 2025IEEE Conference on Computer Communications

This work addresses the deployment challenge of deep reinforcement learning (DRL) in 6G mobile networks stemming from its lack of interpretability. To this end, the study proposes an intent-driven, programmable control framework that uniquely integrates symbolic artificial intelligence with DRL. By leveraging symbolic rules and logical reasoning, the framework generates semantically rich explanations for agent decisions, substantially enhancing the transparency and controllability of learned behaviors. Evaluated on real-world network management tasks, the approach achieves a 12% improvement in median cumulative reward over pure DRL baselines while producing human-understandable decision processes. This integration establishes a novel paradigm for explainable reinforcement learning (XRL), bridging the gap between high-performance learning and interpretable, trustworthy autonomy in complex communication systems.

6 citationsRead paper

On the contraction properties of Sinkhorn semigroups

Mar 12, 2025

This study addresses the exponential convergence of Sinkhorn iterations for general φ-divergences and Kantorovich-type criteria in weighted Banach spaces. Methodologically, it introduces a novel operator semigroup contraction framework grounded in Lyapunov functions, unifying the analysis under minimal regularization conditions. The key contributions are: (1) extending exponential convergence beyond entropy regularization to arbitrary convex φ-divergences and weighted norms; (2) deriving a universal upper bound on the convergence rate and rigorously characterizing its asymptotic behavior toward the Schrödinger bridge solution. The approach integrates Lyapunov stability theory, operator semigroup analysis, and functional-analytic properties of Schrödinger bridges. Empirical validation on linear Gaussian systems and Gaussian mixture models demonstrates superior convergence robustness and computational efficiency in generative modeling tasks.

4 citationsRead paper
Recent publications

Latest Papers