Doctoral School

Reading date: 16/07/2026

• GARCIA RECASENS, POL: Optimizing efficiency and integrity in collaborative AI systems
  
  Author: GARCIA RECASENS, POL
  Programme: DOCTORAL DEGREE IN COMPUTER ARCHITECTURE
  Department: Department of Computer Architecture (DAC)
  Mode: Normal
  Deposit date: 22/06/2026
  Reading date: 16/07/2026
  Reading time: 10:00
  Reading place: C6-E106
  Thesis director: BERRAL GARCÍA, JOSEP LLUÍS | TORRES VIÑALS, JORDI
  Thesis abstract: Modern AI systems are increasingly deployed in collaborative settings where multiple users share data, compute or infrastructure. The design of these systems relies on cooperative assumptions that are routinely violated in practice. These violations degrade both efficiency and integrity across the AI lifecycle, wasting computational resources and corrupting system outputs. For example, participants may not contribute honestly to federated training, memory may be over-allocated in service environments shared by multiple users, and in-context examples used for data generation may be biased or adversarially manipulated.This thesis demonstrates that by designing mechanisms that remain effective when cooperative assumptions fail, the AI lifecycle can be adapted to decentralized deployments while remaining robust to misaligned or adversarial participants. We prove this statement through three contributions that target training, serving, and data generation respectively. First, this thesis proposes FRIDA, a framework that repurposes adversarial privacy attacks to detect free-riders in Federated Learning. Rather than relying on indirect statistical signals that advanced free-riders can mimic, FRIDA directly measures evidence of genuine local training. Our evaluation shows that FRIDA detects adaptive free-riders that evade existing feature-based defenses.Second, this thesis characterizes the throughput-latency frontier of Small Language Model serving and identifies that the Pareto-optimal throughput plateau is reachable within a single accelerator. Through low-level GPU profiling, we show that the plateau is caused by DRAM bandwidth saturation within the attention mechanism, not compute saturation as commonly assumed. Based on this finding, we introduce the Batching Configuration Advisor, which recommends batch sizes and memory allocations that respect latency constraints without over-provisioning, freeing resources for concurrent workloads and achieving throughput improvements of up to 33.7\%.Third, this thesis characterizes adversarial in-context bias propagation as a new attack surface in collaborative LLM-based tabular data generation. We show that statistical biases in prompt examples systematically propagate to generated synthetic data, and that a malicious actor can amplify this effect by injecting feature-aligned examples to target specific subgroups without degrading standard utility metrics. We evaluate mitigation strategies including statistical parity constraints and frequency balancing, showing that while they attenuate bias propagation, the sensitivity of LLMs to adversarial in-context examples remains a persistent challenge.These results demonstrate that explicitly accounting for the failure of cooperative assumptions enables meaningfully more efficient and trustworthy AI systems across training, serving, and data generation.
  

• GOYZUETA MEJÍA, ZARELA MÓNICA: L'orientació de client a partir de l'experiència educativa: la perspectiva d'estudiants de pregrau a les universitats privades del Perú
  
  Author: GOYZUETA MEJÍA, ZARELA MÓNICA
  Programme: DOCTORAL DEGREE IN BUSINESS ADMINISTRATION AND MANAGEMENT
  Department: Department of Management (OE)
  Mode: Normal
  Deposit date: 08/06/2026
  Reading date: pending
  Reading time: pending
  Reading place: pending
  Thesis director: CONSOLACION SEGURA, CAROLINA MARIA | BARREDO IBAÑEZ, DANIEL
  Thesis abstract: Neoliberalism, as a model of global governance, has promoted competition and market efficiency as pillars of economic and social progress. Its influence has permeated various sectors, and higher education has been no exception. In this sphere, this ideology has driven the emergence of marketization, a trend that conceives of the education system under a supply and demand model, where education is understood as a good or service that can be acquired, and whose value is measured primarily by the economic return it generates for the individual and its contribution to the economy as a whole.In this context, the metaphor of the student as customer arises, a concept of business origin that positions the student as the central consumer of an experiential educational service, aimed at providing them with a satisfactory academic foundation for their professional development.This phenomenon has generated extensive debate in literature, with arguments both for and against. However, existing research has not delved deeply into educational experience as a determining variable in shaping the profile of the student as client. Furthermore, most studies have focused on Europe and the United States, leaving Latin America—and particularly Peru—as a largely unexplored context.Within this framework, the purpose of this doctoral thesis is to analyze how client orientation is shaped in undergraduate students at private universities in Peru, considering the dimensions of their educational experience, the influence of demographic variables, and the perspective of academic experts. To this end, a mixed-methods approach was adopted.First, a systematic literature review on the student as customer metaphor (2000–2022) was conducted in the Scopus and Web of Science databases, identifying 83 studies that met the eligibility criteria. Second, the SCOQ (Student Customer Orientation Questionnaire) scale by Koris and Nokelainen (2015) was adapted to the Peruvian context, and 798 surveys were collected from three private universities. The data were analyzed using exploratory factor analysis, which identified the dimensions of the educational experience that define customer orientation in Peruvian students. Subsequently, a multiple regression model was applied to assess the influence of demographic variables. Finally, the results were triangulated with 12 in-depth interviews with academic experts to corroborate and enrich the findings.The research results reveal four main findings. First, the literature is structured around two predominant approaches: a critical perspective, focused on the risks of commodification and the weakening of academic rigor, and a relational perspective, which conceives of the student as a co-producer of the educational service. Second, adapting the SCOQ instrument to the Peruvian context allowed, through exploratory factor analysis, the identification of five factors of the educational experience. In four of these factors, students exhibit a client orientation. In contrast, positive self-demand is the only factor in which this orientation is not observed, highlighting the hybrid and multidimensional nature of the Peruvian student, who combines service demands with academic responsibility. Third, the multiple regression analysis, while having limited explanatory power, showed that some demographic variables have predictive value, with years of study standing out as the most influential factor. Finally, the triangulation with experts showed that teachers adopt a conditional acceptance stance, in which they recognize certain expectations inherent in a service logic on the part of students, but subordinate their validity to the preservation of academic rigor and teacher autonomy.
  

• JADOON, MUHAMMAD AWAIS: Machine Learning-based Random Access Techniques for Massive Connectivity
  
  Author: JADOON, MUHAMMAD AWAIS
  Programme: DOCTORAL DEGREE IN SIGNAL THEORY AND COMMUNICATIONS
  Department: Department of Signal Theory and Communications (TSC)
  Mode: Normal
  Deposit date: 08/06/2026
  Reading date: 16/07/2026
  Reading time: 11:30
  Reading place: Auditorium B6 Building, Av. Carl Friedrich Gauss, 11, PMT, Castelldefels (Barcelona)
  Thesis director: NAVARRO RODERO, MÒNICA | PASTORE, ADRIANO
  Thesis abstract: Massive machine-type communication (mMTC) underpins the 5G-and-beyond vision of supporting ultra-dense networks of low-complexity, low-power, battery-operated devices whose traffic characteristics differ significantly from traditional human-type communication, being more sporadic, uncoordinated, and often event-driven. Managing medium access for such a massive number of devices is a central challenge in mMTC. Traditional scheduling or grant-free Random access (RA) approaches often incur excessive signaling overhead, high collision rates, and offer limited adaptability in environments where traffic patterns vary. This necessitates the design of novel medium access schemes that can scale to massive numbers of devices.In this thesis, we employ multi-agent reinforcement learning (MARL) to design distributed grant-free RA policies that accommodate dynamic traffic conditions and provide service to large populations of devices. We first develop a single-channel environment where the devices learn transmission policy using one-bit broadcast feedback and local packet buffer states using the Deep Q-Learning (DQN) algorithm. Although this approach surpasses baseline Exponential Backoff (EB) techniques in terms of throughput and fairness under regular (Poisson) traffic, it relies on a single-agent training framework and does not fully exploit the Centralized Training and Decentralized Execution (CTDE) principle. Recognizing the importance of scalability and effective coordination in large networks, we next adopt advanced MARL algorithms—Value Decomposition Networks (VDN) and QMIX, in which a global Q-value is computed from individual Q-values, thus leveraging global network information during training.To evaluate fairness, we introduce the age of packet (AoP) metric, which quantifies the staleness of untransmitted packets in device buffers. We then extend our models to accommodate bursty and correlated traffic arrivals, demonstrating that the proposed MARL schemes can adapt effectively to sudden device activation or shifting arrival patterns. We also investigate how user identification impacts policy learning and fairness, as devices may dynamically join or leave the network. We show how these design choices can influence fairness and allow devices to leave/join the network. Additionally, we broaden the scope beyond a single channel, incorporating multiple orthogonal resources to highlight the generality of these approaches. Throughout our extensive simulations, all MARL-based methods consistently outperform EB schemes. Our findings underscore the promise of MARL-driven solutions for mMTC and future wireless systems.
  

• MARTÍ FLORENCES, MIQUEL: Regression-Based Estimation of Internal Variables in Lithium-Ion Batteries
  
  Author: MARTÍ FLORENCES, MIQUEL
  Programme: DOCTORAL DEGREE IN AUTOMATIC CONTROL, ROBOTICS AND VISION
  Department: Department of Automatic Control (ESAII)
  Mode: Normal
  Deposit date: 04/06/2026
  Reading date: 16/07/2026
  Reading time: 16:00
  Reading place: Sala d'Actes de la Facultat de Matemàtiques i Estadística (FME), Campus Diagonal Sud, Carrer de Pau Gargallo, 14, 08028 Barcelona
  Thesis director: COSTA CASTELLO, RAMON | CECILIA PIÑOL, ANDREU
  Thesis abstract: The decarbonisation of energy systems, the increasing penetration of renewable generation, and the electrification of transport have made lithium-ion batteries a key enabling technology in modern power applications. Their versatility and high energy density explain their central role in electric vehicles and stationary energy storage. However, their safe and efficient operation depends on battery-management systems capable of estimating internal quantities that cannot be measured directly, such as state of charge, open-circuit voltage, and relevant electrical parameters. This task is difficult because battery behaviour is nonlinear, operating conditions are variable, and the measurements normally available in practice are limited.This thesis addresses these challenges through a practical estimation framework for real-time battery-management applications. It reviews the main modelling and estimation approaches for lithium-ion batteries and justifies the use of low-order equivalent circuit models as a suitable compromise between physical relevance and online implementability. It also reviews the estimation-theoretic basis required to work under realistic operating conditions, with particular attention to observability and excitation. A central point is that classical recursive estimators often rely on persistent excitation, whereas battery current profiles are usually dictated by operation and may only be informative over finite intervals. To bridge this gap, the thesis adopts novel regression-based tools such as the generalised parameter estimation based observer, dynamic regressor extension and mixing, and non-asymptotic least squares estimators.On this basis, two application-oriented estimation architectures are developed. The first jointly estimates the dominant parameters of a first-order equivalent circuit model and the instantaneous open-circuit voltage from online current and terminal-voltage measurements, and is validated through simulations and experiments. The second extends the plug-and-play perspective to the online identification of parametrised open-circuit-voltage/state-of-charge relations from finite informative intervals. Overall, the thesis shows that useful battery monitoring can be achieved without extensive offline parametrisation campaigns or artificially persistent excitation, thereby supporting more practical battery-management solutions.
  

• RHOUMA, ALI: Operationalizing the Water–Energy–Food–Ecosystems Nexus for the Sustainability Assessment of Mediterranean Farming Systems
  
  Author: RHOUMA, ALI
  Programme: DOCTORAL DEGREE IN SUSTAINABILITY
  Department: University Research Institute for Sustainability Science and Technology (IS.UPC)
  Mode: Normal
  Deposit date: 24/04/2026
  Reading date: 16/07/2026
  Reading time: 12:30
  Reading place: Salón de Grados de la EEABB Campus of Castelldefels
  Thesis director: GIL ROIG, JOSE MARIA | BROUWER, FLOOR
  Thesis abstract: Agricultural systems face increasing pressures from water scarcity, climate change, and environmental degradation, particularly in the Mediterranean region where resource constraints are intensifying. Addressing these interconnected challenges requires integrated analytical approaches capable of capturing the complex interactions between water, energy, food production, and ecosystems. The Water–Energy–Food–Ecosystems Nexus has emerged as a promising framework for supporting sustainable resource management; however, its operationalization at the farm level remains limited. This thesis aims to advance the application of the WEFE Nexus for the sustainability assessment of farming systems by developing an integrated analytical framework and practical evaluation tools.The research adopts a progressive methodological approach. First, water-related sustainability indicators specifically the water footprint and water scarcity footprint are applied to assess the pressures of agricultural production on water resources. These indicators provide an initial understanding of resource use efficiency and highlight critical water-related challenges in farming systems. Building on this foundation, the thesis develops a WEFE Nexus assessment framework based on system dynamics modelling to capture interactions between water use, energy consumption, food production, and ecosystem impacts. The framework is implemented through a user-friendly decision-support tool designed to support sustainability assessments at the farm level.The developed framework is subsequently applied to evaluate WEFE Nexus solutions in agricultural systems, with a particular focus on agroecological practices. By integrating multiple indicators related to resource efficiency, environmental performance, and agricultural productivity, the analysis explores how agroecological approaches influence the performance of farming systems within the WEFE Nexus. The results demonstrate that agroecological practices can improve resource use efficiency, reduce environmental pressures, and enhance the resilience of farming systems.Finally, the outcomes of the WEFE Nexus assessment are translated into key performance indicators linked to the Sustainable Development Goals (SDGs). This step enables the quantification of how WEFE-based agricultural practices contribute to broader sustainability objectives and global development targets. By linking farm-level sustainability assessment with the SDG framework, this research provides a novel methodological contribution for evaluating the sustainability impacts of agricultural practices.This thesis contributes to bridging the gap between WEFE Nexus theory and practical agricultural sustainability assessment. The proposed framework offers a robust approach for evaluating sustainable farming systems and provides valuable insights for policymakers, researchers, and stakeholders seeking to promote resilient and resource-efficient agriculture in the Mediterranean region and beyond.