Màster Oficial - Fonaments de la Ciència de Dades
URI permanent per a aquesta col·leccióhttps://hdl.handle.net/2445/133281
Treballs finals del Màster de Fonaments de Ciència de Dades de la Facultat de Matemàtiques i Informàtica de la Universitat de Barcelona
Examinar
Enviaments recents
Mostrant 1 - 20 de 122
Treball de fi de màster
Training Small Language Model Rankers with Supervised and Reinforcement Learning(2026-06-01) Virgili Mateu, Júlia; Vitrià i Marca, JordiLarge language models make it possible to express recommendation as a constrained generation problem: given a user history and a candidate set, the model writes an ordered list rather than scoring each item separately. This interface is flexible, but relying on frontier-scale models is not always compatible with local deployment, cost, latency, privacy, or energy constraints. Small language models therefore offer an important test case: can a compact ranker be aligned well enough for a structured recommendation task? This thesis investigates when supervised fine-tuning is sufficient and when reinforcement learning adds useful signal for such a small language model ranker. The study uses Qwen3-4B in a fixed local-identifier interface and compares imitation learning with two reward-based training strategies: one gives credit to the generated list as a whole, while the other tries to distribute credit according to the positions in the ranking. Because generated rankings can be parsed, validated, and scored automatically, recommendation provides a natural setting for studying reinforcement learning from verifiable rewards. The evaluation also includes simple classical candidate-set baselines, so the generative rankers can be interpreted against lightweight score-and-sort methods rather than only against each other. The experiments show that method suitability depends on the information structure of the task. In a binary setting, supervised fine-tuning is already well matched to the objective, while position-specific credit has limited signal to exploit. In a graded setting, reinforcement learning becomes more meaningful because rewards express gradual, position-sensitive utility; both reward-based variants improve the head of the generated ranking relative to the supervised baseline and substantially outperform the strongest simple candidate-set baseline considered. The thesis therefore argues for choosing alignment methods according to binary relevance, graded relevance, position sensitivity, and reward information density, rather than treating reinforcement learning as a generic improvement over imitation.Treball de fi de màster
Predicting venous thromboembolism in cancer patients using contrastive learning(2026-06-27) Villaró Cañizal, Júlia; Lobato Delgado, Bárbara; Pujol Vila, OriolThis thesis studies whether contrastive learning can improve the prediction of venous thromboembolism (VTE) in cancer patients using a particular dataset, which contains very few samples and a large number of clinical variables. In this setting, standard supervised learning models often struggle to capture the complex patterns required for accurate prediction. To address these limitations, the work investigates whether contrastive learning can learn more robust patient representations from limited data. It’s also examined whether adding a feature gating mechanism to the contrastive model can down-weight less informative variables, which would improve model interpretability.Treball de fi de màster
What Matters in Image-based Cloth Size Estimation?(2026-06-28) Teker, Izel; Escalera Guerrero, SergioEstimating garment measurements from a single photograph is a supervised learning problem starved of supervision: real sizing charts are produced by hand and exist only at the scale of a few hundred garments, while the synthetic corpora large enough to train deep models carry no measurement labels. This thesis connects the two. I build an automated pipeline1 that derives physical measurements directly from the 3D garment meshes of SewFactory, producing a large-scale synthetic measurement dataset aligned to the 19-dimensional MV-Fashion schema and combinable with the real corpus. Using this hybrid data, I ask what matters for image-based garment size estimation, and isolate three factors. First, data composition is the dominant controllable factor: hybrid synthetic–real training improves on a real-only baseline, most strongly in the garment groups where real annotation is scarcest. This holds despite measurable noise in the synthetic labels, and a controlled probe shows the residual synthetic–real gap to be one of label quality and coverage, not image appearance. Second, a pose-canonicalisation stage inherited from sewing-pattern reconstruction does not transfer to measurement: added as a frozen pixel-fidelity pre-processor it degrades accuracy, because it smooths away the free-raping geometry the measurements depend on. Third, for the hardest dimensions the binding constraint is the single near-frontal view itself, which no amount of data can overcome. The broader contribution is methodological: for this task synthetic data is worth having for its labels and coverage rather than as an appearance bridge, and a preprocessor optimised for one task does not transfer for free to another whose objective differs.Treball de fi de màster
Analyzing Data-Efficient Reinforcement Learning for LLM-Based Recommendation through Data Selection(2026-06-28) Soler Sales, Arnau; Seguí Mesquida, SantiLarge language models can be trained as recommenders that optimize ranking quality directly, but training them with reinforcement learning over a full dataset is costly. This motivates data selection: choosing a small, high-value subset to train on. This thesis studies MiniRec (L. Wang et al., 2026), a method that scores each example for learnability, representativeness, and diversity using a small proxy model, and is presented as a general, model-agnostic approach to data-efficient RL-based recommendation. We first reproduce it on Amazon CD/Vinyl, where a 256-sample subset recovers 93.5% of full-data NDCG@10 at roughly a tenth of the training cost, matching the original paper. We then transfer the unmodified method to two different settings, MIND news and Yelp business ranking (Wu et al., 2020; Yelp Inc., 2023), where it scores below random selection on both. A signal-by-signal analysis explains why: on MIND, the reward signal is almost entirely explained by two unrelated factors, the number of candidates and the position of the correct answer, while the other two signals carry no information at all; Yelp shows the same pattern, even though an independent encoder finds real differences between the same items, which places the fault in the proxy rather than in the data. Every method we test on these two datasets, including MiniRec itself, ties random selection. This is not because the choice of examples is irrelevant: a selector built only from information outside the proxy, using simple facts about user history and business category, performs slightly better than random. The contribution of this thesis is this diagnosis: because all three of MiniRec’s signals are read from a single frozen proxy, its reported generality does not survive a genuine change of task, and the usable signal, where it exists, lies outside the proxy rather than within it.Treball de fi de màster
Multidimensional Student Profiling for Understanding Educational Trajectories and Intervention Impact: The Pentabilities Example(2026-06-28) Segura Lladó, Lluc; Calsamiglia, Caterina; Puertas i Prats, EloiThis thesis analyses the value of students’ baseline sociodemographic and socioemotional characteristics in understanding their educational trajectories and assessing the impact of the Pentabilities programme in high-complexity secondary schools in Catalunya. Using data from a randomised controlled trial, the results of the treatment and control groups are compared in terms of final marks, absenteeism and change in pupils’ perception of the teacher, and predictive models and clustering techniques are applied to identify student profiles. The results do not show a clear average effect of the intervention on the outcomes analysed. However, socioemotional variables provide relevant information for predicting low academic achievement, particularly when combined with sociodemographic variables. Furthermore, cluster analysis enables the identification of student profiles with interpretable differences in social and emotional terms, and their connection with academic outcomes. Overall, the results highlight the potential of socioemotional information to better understand student heterogeneity and support more accurate identification of educational trajectories at risk.Treball de fi de màster
Causal Machine Learning for Markdown Price Optimization in Retail(2026-06-27) Sanfeliu Ichihashi, Joel; Pujol Vila, Oriol; Gutiérrez Meyer, DanielThis thesis develops a causal machine learning framework for markdown price optimization in fashion retail. The objective is to improve the elasticity curves used by the existing markdown optimizer, which under a purely predictive approach often recommends the minimum allowed discount instead of identifying a balanced markdown level. This is problematic because markdown pricing should balance the increase in demand from deeper discounts against the loss in unit price. Using historical sales and pricing data from Mango, the study applies Double Machine Learning to control for pre-pricing factors such as stock, previous sales, product characteristics, and calendar timing. A Causal Forest is then used to estimate heterogeneous price effects, allowing the demand response to vary across products and moments of the sales campaign. The final model is selected after comparing alternative causal specifications and is chosen for its ability to generate smoother and more economically reasonable markdown curves. The results show that the causal approach maintains predictive performance while improving the quality of the simulated curves used by the optimizer. The proposed framework therefore improves the curve-generation step while remaining compatible with the existing optimization pipeline.Treball de fi de màster
Spatio-temporal aware streamflow prediction in river basins using Graph Neural Networks(2026-06-28) Raso Alonso, Carlos; Juez, Carmelo; Vitrià i Marca, JordiThis thesis proposes a model for streamflow prediction in river basins that combines Graph Neural Networks (GNNs) with Long Short-Term Memory (LSTM) cells. The model is trained and evaluated with data from gauging stations located in the Aragon River basin, which forms part of the northern Ebro basin. LSTM models have become prominent in recent hydrological modelling, but they often underpredict peak streamflow events, possibly due to their limited spatial awareness of river basin topology. GNNs provide a natural framework for encoding spatial and hydrological relationships between gauging stations. The proposed model therefore aims to address the spatial limitations of LSTM-based approaches by integrating temporal modelling and graph structure within a single framework. The results show that the proposed GNN-LSTM model outperforms the reference LSTM model in terms of Nash–Sutcliffe efficiency (NSE). Moreover, for some gauging stations, it also improves performance during flood events with return periods greater than two years. These findings suggest that graph-based temporal models can enhance LSTM-based streamflow prediction and may be extended to larger river basins. In addition, the proposed framework enables streamflow estimation at ungauged locations within the basin.Treball de fi de màster
Improving Machine Learning Models for Fluorescence-Based Classification of Airborne Bacteria(2026-06-28) Perry, Christopher; Fontal Navarrete, Alejandro; Rodó, XavierRapid identification of airborne bacteria is critical for environmental monitoring, public health, and biodefense, yet conventional laboratory methods are unsuitable for continuous real-time analysis. This thesis investigates machine learning and deep learning approaches for bacterial bioaerosol classification using Laser-Induced Fluorescence (LIF) measurements acquired with a modified Rapid-E instrument. Six progressively more advanced models were developed, progressing from a reproduced Random Forest baseline to a robust multimodal convolutional neural network integrating fluorescence spectra, fluorescence lifetime, scattering images, and particle-level metadata. The proposed model achieved a test accuracy of 88.91%, a balanced accuracy of 79.36%, and a macro F1-score of 0.787, outperforming both the reproduced Random Forest baseline and the previously published Rapid-E model. Finally, a real-time software framework was developed to automate data processing, model inference, prediction storage, and live visualization, demonstrating the feasibility of deploying multimodal deep learning for continuous bacterial bioaerosol monitoring.Treball de fi de màster
Efficient Adaptation of Vision-Language Models for Cloud Image Classification Using LoRA(2026-06-28) Morera Barrios, Ignacio Javier; Vitrià i Marca, JordiAccurate cloud classification from ground-based images matters for short-range weather forecasting, climate monitoring, and solar irradiance prediction. In practice, it is a hard visual problem: cloud types share overlapping textures and shapes, their appearance changes with lighting and altitude, and even trained meteorologists sometimes disagree on the correct label. A practical barrier is that available datasets were collected and labelled independently using different taxonomies, so combining them for training requires reconciling incompatible class definitions. further constraint is that achieving high accuracy with large pre-trained models can require computational resources that are unavailable in many practical scenarios. This thesis addresses both issues. First, we build a merged benchmark from four public datasets: SWIMCAT, CCSN, FabraClouds256, and FabraSwimcat, unified under a shared eleven-class taxonomy. Second, we adapt Qwen3.5-0.8B, a compact vision-language model, to this benchmark in two stages: supervised fine-tuning (SFT) with single-label supervision to learn the cloud domain and vocabulary, followed by reinforcement learning via Group Relative Policy Optimisation (GRPO), which trains the model to produce a ranked list of all eleven classes and rewards it for placing the correct class at the top. Low-Rank Adaptation (LoRA) keeps trainable parameters at just 0.13% of the backbone, so the full pipeline fits on free-tier cloud GPUs.Treball de fi de màster
Federated Learning and Model Copying in Machine Learning: A Proof of Concept(2026-06-28) Martín Delgado, María; Pujol Vila, Oriol; Rovira Esteva, MurielThis thesis introduces a new Federated Learning approach using copies of Machine Learning classifiers. Increasing concerns about data privacy motivated the creation of Federated Learning algorithms that allowed Machine Learning classifiers to be trained collaboratively among different clients without sharing their data. In this context, we propose a new Federated Learning strategy using copies that deals with one of the biggest problems of the traditional Federated Learning approach: model heterogeneity among participants. Copies have the ability to reproduce the decision behaviour of arbitrary classifiers independently of their architecture and to leverage their knowledge. Since they are also data-agnostic and need no real data to be trained, copies make ideal candidates to address model heterogeneity among clients while preserving the data privacy principle that is fundamental to Federated Learning. We describe in detail the integration of copies in Federated Learning systems, show how we have implemented it, and test its feasibility in two experiments of different nature. Finally, we include the results of the experiments, which verify the efficacy of proposed strategy, and present some ideas that could extend the analysis done in this project.Treball de fi de màster
Topological Data Analysis in Clinical Trial Data: A study of Treatment Retention in Opioid Use Disorder(2026-06-28) Llauger, Miquel Àngel; Drach, KostiantynThe Mapper algorithm of Topological Data Analysis (TDA) captures the shape of a point cloud as a graph, but in doing so it collapses many data points into each node. This thesis introduces the Metric Graph, a variant of Mapper in which every point is retained as a vertex and a projection cover is used only to constrain the edges of an underlying $\varepsilon$-neighbourhood graph in a given metric space. Because each vertex now corresponds to a single subject, we enable node-level prediction via Graph Neural Networks (GNNs). We apply the method to the National Institute on Drug Abuse (NIDA) CTN-0051 Extended-Release Naltrexone vs. Buprenorphine-Naloxone (X:BOT) study for Opioid Use Disorder (OUD), focusing on patient retention. We build Metric Graphs at trial weeks 8 and 12 using baseline and engagement features and train Graph Convolutional Networks (GCNs) to predict dropout, benchmarking against non-graph classifiers. We report the experimental results, discuss the parameter choice and assess whether GNNs outperform classical models when they operate on the Metric Graph. Datasets, code and details publicly available at GitHub.Treball de fi de màster
Learning sweep embeddings from blind obstetric ultrasound for gestational age prediction(2026-06-28) Hristov, Christin; Dang, Vien NgocGestational age (GA) prediction from blind-sweep obstetric ultrasound could help extend fetal dating to settings with limited access to trained sonographers. However, GA labels for training predictive models are harder to obtain than unlabeled sweep videos. This thesis studies whether a frozen self-supervised ultrasound encoder, combined with a lightweight sweep-level aggregator, can support label-efficient GA prediction when only a subset of training sweeps has GA labels. Using 1,314 blind-sweep videos from two acquisition sites, Spain and Kenya, three frame-encoder pretraining conditions and four sweep aggregators are compared. The resulting sweep embeddings are evaluated for GA prediction under full and reduced label budgets. Encoder pretraining quality appears to be the dominant factor among those tested. The best-performing encoder, which combines general ultrasound pretraining with continued pretraining on in-domain sweep data, outperforms a from-scratch encoder by 5.1 days mean absolute error. Aggregator choice and frozen versus trainable aggregation, by contrast, show much smaller effects, spanning only 1.14 days once the encoder is fixed, with frozen aggregators reaching similar performance to aggregators trained jointly with the prediction head. This encoder advantage also appears to carry over to label efficiency: continued pretraining at only 10% of training labels already reaches a weaker encoder’s accuracy at 100% of labels. Aggregator choice and training regime remain secondary across the label budgets tested. A gap of approximately 5 days remains to a fully supervised reference at full label availability, which may be associated with keeping the encoder frozen. For this dataset and fixed frame-selection pipeline, the results suggest prioritizing encoder pretraining over aggregator complexity, since aggregators pretrained without GA labels showed no consistent MAE drop compared with trainable aggregators.Treball de fi de màster
Automated Regulatory Compliance and Readability Assessment of Clinical Informed Consent Documents Using Local Large Language Models(2026-06-28) Hernández Castilla, Ana Valentina; Häyrinen, Noora; Modamio Charles, Pilar; Puertas i Prats, EloiInformed Consent Forms (ICF) and Patient Information Sheets (PIS) are central to clinical research: they guarantee that participants enter a trial having received a complete and intelligible description of its purpose, procedures, risks and rights. Their regulatory review, however, remains an almost entirely manual task: a reviewer must verify, item by item, that dozens of required elements are present in a document that often exceeds twenty pages, and must do so for every protocol amendment, a formal modification to the trial design after approval that requires a full re-review of the updated documents. This thesis presents a modular pipeline, which is a sequence of independent, replaceable processing stages communicating through file-based interfaces, that automates this review using locally hosted Large Language Models, so that no document ever leaves the user’s machine. The system, designed to operate on documents in either Spanish or English, extracts structured content from PDF documents (including scanned, vector-encoded and partially redacted pages), heuristically separates the PIS and ICF sections, evaluates each document against a 62-item regulatory checklist (49 PIS items + 13 ICF items) using a local LLM (Qwen2.5-14B served through Ollama), and produces a readability and clarity analysis for the patient-facing language. Results are delivered as multi-sheet Excel, Word and PDF reports. A separate validation module compares the system’s output against a manual expert evaluation of twenty documents, computing Cohen’s Kappa and per-item agreement metrics. The pipeline was tested on a corpus of 41 Spanish-language clinical documents drawn from ten therapeutic areas. Three additional experiments compared model sizes (7B vs 14B), assessed module independence, and tested a prompt modification. The work demonstrates that local LLMs can deliver a defensible, reproducible and confidentiality-preserving compliance assistant for a domain where commercial cloud APIs are not an option. The full source code is available at https://github.com/AnaVHdez/thesis_Final.Treball de fi de màster
Logic Tensor Networks for Knowledge-Infused Learning: Foundations, Experiments, and Applications(2026-06-27) Guiot Cusidó, Quim; Pujol Vila, OriolThis thesis investigates Logic Tensor Networks (LTN) as a neurosymbolic framework for knowledge infused learning, focusing on their theoretical underpinnings, practical behavior, and applicability to real world tasks. We first present a self contained exposition of Real Logic, detailing how symbols are grounded into tensors, how fuzzy logical operators are instantiated in a differentiable setting, and how learning is framed as maximizing the satisfiability of a Real Logic theory. Building on this foundation, we design a controlled 1D regression problem where a neural model predicts a person’s weight from their height under different logical configurations. Through three experiments (baseline regression, a favorable monotonicity axiom, and a deliberately misleading monotonicity axiom) we study how prior knowledge encoded as axioms affects in distribution performance, out of distribution generalization, and the truth degree of queries. The results show that well aligned axioms can improve both regression accuracy and robustness, while incorrect axioms systematically degrade performance, highlighting LTN as a sensitive interface between data and prior knowledge. Finally, we explore a real life recommender system case study, where LTN are used to mitigate data sparsity by injecting genre based user preferences into a matrix factorization pipeline. Altogether, this work illustrates both the promise and the risks of knowledge infused learning with LTN. The following link will take the lector to the GitHub Repository containing the performed and developed experiments discussed on chapter 4: My GitHub Repository. URL to copy paste on your browser: https://github.com/srguiot14/Logic-Tensor-Networks-for-Knowledge-Infused-Learning-Foundations-Experiments-and-Applications.Treball de fi de màster
Algorithmic Fairness in ICU Mortality Prediction: A Multi-Model Analysis of Racial Disparities, Structural Bias, and Mitigation Strategies Using MIMIC-IV(2026-06-28) Grgurevic, Ino; Igual Muñoz, LauraMachine learning models are increasingly used to predict in-hospital mortality for intensive care patients, which raises the question of whether they perform equally well for every population subgroup. This work uses MIMIC-IV a large and freely available critical care database to compare four model architectures. Two of these models, XGBoost and the deep feedforward network, use static summary features computed over the first twenty-four hours, while the LSTM and Transformer process the full hourly time series instead, and studies whether their predictions show systematic performance gaps across racial groups, what causes any such gaps, and whether they can be reduced. Across all four models the same small racial gap emerges, modest in size and consistent regardless of architecture. A series of experiments locates the source of this gap in the data itself rather than in any model, and shows that it responds only to changes at the level of data collection, not to adjustments made to the model afterward. The findings point to more equitable clinical measurement and data collection as the most reliable route to fairer ICU prediction.Treball de fi de màster
Precise Action Spotting in untrimmed soccer videos(2026-06-28) García Arias, Javier; Escalera Guerrero, Sergio; Clapés i Sintes, AlbertAction detection involves locating the precise moment when each action occurs in a long, untrimmed video. Most prior work on football action spotting has been conducted on broadcast footage; we instead study the task on the GameOn dataset, which provides panoramic camera coverage and synchronised player and ball tracking, but introduces several challenges: actions are densely annotated and temporally close to one another, the panoramic view provides comparatively weak visual evidence, and combining the visual and tracking modalities without destabilising training is a non-trivial design problem. This thesis addresses these challenges on two fronts. First, we conduct a systematic study of temporal modeling architectures for spotting, comparing single-scale recurrent, convolutional, and attention-based modules and proposing a multi-scale Temporal Encoder–Decoder (TED) module that integrates long-range context while preserving frame-level resolution. Second, we study the integration of the tracking modality through a Graph Isomorphism Network and identify intermediate auxiliary supervision over both modality branches as a structural prerequisite for the multimodal model to outperform either modality in isolation. The proposed multimodal architecture reaches 75.27 aMAP@1 and 77.35 aMAP@2 on the held-out test split, improving on the vision-only and tracking-only baselines by +6.42 and +7.22 aMAP@1 respectively. The full implementation is publicly available.Treball de fi de màster
Aligning Credit Scoring with the EU AI Act: Bias Mitigation Techniques in Financial Machine Learning(2026-06-28) Esbert Barber, Daniel; Vitrià i Marca, JordiThe deployment of machine learning algorithms in retail banking has transformed credit risk assessment. However, proprietary, opaque predictive systems raise significant concerns regarding transparency, algorithmic bias, and regulatory compliance. This work aims to dissect the implications the EU AI Act will have on credit scoring models and analyze the trade-offs between profitability and fairness that take place when applying different bias mitigation techniques in a realistic credit scoring setting.Treball de fi de màster
Fairness-Aware Multimodal Machine Unlearning(2026-06-28) Dimitrova, Ralitsa Ivanova; Jacques Junior, Julio C. S.We evaluate fairness in protected demographic groups under sample-level and featurelevel unlearning. Our results show that fairness is not consistently aligned with utility, thus fairness metrics should be incorporated into unlearning evlauation. We compare impact of random and targeted forget sets on fairness across demographic subgroups and demonstrate that analysing only random forgetting may hide potential fairness degradation for those subgroups, which on the other hand are exposed by a specially crafted forget set. We analyse feature revocation impact on fairness across demographic subgroups, and our results suggest that impact depends on the predictive relevance of the removed attribute and the underlying dataset distribution, with demographic feature removal not necessarily reducing performance. These findings highlight that feature-level unlearning performance likely depends on prior feature importance and dataset structure, making feature analysis an essential component when evaluating unlearning in this setting. Overall, we show that utility alone is insufficient to characterise machine unlearning behaviour and may not capture subgroup-specific effects across protected demographic groups.Treball de fi de màster
Comprehensive Format Benchmark for Recommender SystemWorkloads(2026-06-27) Cruz Pérez, Javier de; Cumplido Domínguez, Raúl; Puertas i Prats, EloiChoosing a file format for training data is usually a routine technical decision based on experience or common practice, without much understanding of how it affects later stages of the process. This thesis explores those effects for implicit-feedback recommender systems by comparing five file formats: CSV, Parquet, ORC, Arrow IPC, and Lance, across the entire workflow, from data storage to the quality of the final trained model, using $\sim 15,\mathrm{M}$ MovieLens ratings obtained from the original file with $\sim 32,\mathrm{M}$ rating. Three studies are presented. The first evaluates the data-loading lifecycle by breaking read time into raw I/O, decoding, and materialization stages, and measures the fraction of time attributable to file format, denoted by $f$, across four recommender models with increasing computational cost. The second develops a predictive cost model of the form $T_{read}=\alpha_f+\beta_1 g N+\beta_2 g N \log_2 N$ using nine calibration dataset sizes and validates it on an out-of-core dataset with synthetic and real data. The third study examines how file format choice affects recommendation quality by linking it directly to the NDCG@10 metric. Two key findings emerge. The first is a ranking inversion: the format that is best for storage efficiency (ORC+zstd, which is $\sim 5\times$ smaller than CSV) is not the fastest format for reading data. The second is a regime flip: when the entire dataset fits in memory, the format-attributable fraction $f$ approaches zero as the number of training epochs increases, meaning that file format has little impact on long training runs. In contrast, when training is performed out-of-core, $f$ remains constant, so file format continues to have a significant effect regardless of the number of epochs. These effects also extend to recommendation quality. The time required to reach 95% of the maximum NDCG@10 is an affine function of read time, with a slope of $E_q=3$, which was verified exactly. As a result, the ranking inversion observed in runtime also appears when considering quality over time. These results lead directly to practical recommendations. Arrow IPC is the preferred choice for repeated out-of-core training, ORC is best suited for long-term archival storage, and any columnar format is preferable to CSV when a dataset is read more than five times. The predictive model also makes these recommendations transferable across different hardware platforms, requiring only three parameters per format to be recalibrated. The notebooks used in this project can be found here https://github.com/decruz26/TFM.Treball de fi de grau
Tower Parrot: Learning Air Traffic Conflict Resolution from Human Controllers via Adversarial Inverse Reinforcement Learning(2026-06-01) Dalmau Codina, Ferran; Cos Aguilera, Ignasi; Dalmau Codina, Ramon[eng] Air traffic controllers resolve thousands of potential mid-air conflicts daily, yet automation research has struggled to replicate these decisions. Reinforcement learning approaches have consistently failed because their action spaces (heading and speed changes) do not reflect controller practice, and their hand-crafted reward functions cannot capture the complexity of expert decision-making. We present Tower Parrot, the first application of Inverse Reinforcement Learning (IRL) to air traffic conflict resolution, learning the reward function directly from data rather than engineering it manually. Our approach combines Adversarial IRL (AIRL) with a pointer-network policy that selects among flight-plan waypoints mirroring the DIRECT-TO instructions controllers actually issue. Building on the publicly available Bordeaux ACC deconfliction dataset (olive2024deconflictions), we contribute a two-stage instruction-recovery algorithm that extracts 8,800 labeled demonstrations from 31,000 encounters, exploiting structure not identified in prior work. The central finding is that safety constraints emerge from pure imitation. The ICAO-mandated 5 nautical mile separation minimum is encoded nowhere in our system, yet the AIRL policy maintains separation above this threshold in 100% of 200 held-out encounters, whereas behavioral cloning violates it in 27% of cases. Beyond regulatory compliance, the learned policy intervenes earlier, achieves larger separation margins, and produces smoother trajectories, reflecting implicit controller objectives recovered from data rather than imposed through manual tuning. These results demonstrate that operationally safe, human-aligned conflict resolution can be learned end-to-end from historical data, establishing a new paradigm for decision-support in aviation’s most safety-critical domain.