Universität Stuttgart
Permanent URI for this communityhttps://elib.uni-stuttgart.de/handle/11682/1
Browse
44 results
Search Results
Item Open Access Improving usability of gaze and voice based text entry systems(2023) Sengupta, Korok; Staab, Steffen (Prof. Dr.)Item Open Access Models for data-efficient reinforcement learning on real-world applications(2021) Dörr, Andreas; Toussaint, Marc (Prof. Dr.)Large-scale deep Reinforcement Learning is strongly contributing to many recently published success stories of Artificial Intelligence. These techniques enabled computer systems to autonomously learn and master challenging problems, such as playing the game of Go or complex strategy games such as Star-Craft on human levels or above. Naturally, the question arises which problems could be addressed with these Reinforcement Learning technologies in industrial applications. So far, machine learning technologies based on (semi-)supervised learning create the most visible impact in industrial applications. For example, image, video or text understanding are primarily dominated by models trained and derived autonomously from large-scale data sets with modern (deep) machine learning methods. Reinforcement Learning, on the opposite side, however, deals with temporal decision-making problems and is much less commonly found in the industrial context. In these problems, current decisions and actions inevitably influence the outcome and success of a process much further down the road. This work strives to address some of the core problems, which prevent the effective use of Reinforcement Learning in industrial settings. Autonomous learning of new skills is always guided by existing priors that allow for generalization from previous experience. In some scenarios, non-existing or uninformative prior knowledge can be mitigated by vast amounts of experience for a particular task at hand. Typical industrial processes are, however, operated in very restricted, tightly calibrated operating points. Exploring the space of possible actions or changes to the process naively on the search for improved performance tends to be costly or even prohibitively dangerous. Therefore, one reoccurring subject throughout this work is the emergence of priors and model structures that allow for efficient use of all available experience data. A promising direction is Model-Based Reinforcement Learning, which is explored in the first part of this work. This part derives an automatic tuning method for one of themostcommonindustrial control architectures, the PID controller. By leveraging all available data about the system’s behavior in learning a system dynamics model, the derived method can efficiently tune these controllers from scratch. Although we can easily incorporate all data into dynamics models, real systems expose additional problems to the dynamics modeling and learning task. Characteristics such as non-Gaussian noise, latent states, feedback control or non-i.i.d. data regularly prevent using off-the-shelf modeling tools. Therefore, the second part of this work is concerned with the derivation of modeling solutions that are particularly suited for the reinforcement learning problem. Despite the predominant focus on model-based reinforcement learning as a promising, data-efficient learning tool, this work’s final part revisits model assumptions in a separate branch of reinforcement learning algorithms. Again, generalization and, therefore, efficient learning in model-based methods is primarily driven by the incorporated model assumptions (e.g., smooth dynamics), which real, discontinuous processes might heavily violate. To this end, a model-free reinforcement learning is presented that carefully reintroduces prior model structure to facilitate efficient learning without the need for strong dynamic model priors. The methods and solutions proposed in this work are grounded in the challenges experienced when operating with real-world hardware systems. With applications on a humanoid upper-body robot or an autonomous model race car, the proposed methods are demonstrated to successfully model and master their complex behavior.Item Open Access Concepts and methods for the design, configuration and selection of machine learning solutions in manufacturing(2021) Villanueva Zacarias, Alejandro Gabriel; Mitschang, Bernhard (Prof. Dr.-Ing. habil.)The application of Machine Learning (ML) techniques and methods is common practice in manufacturing companies. They assign teams to the development of ML solutions to support individual use cases. This dissertation refers as ML solution to the set of software components and learning algorithms to deliver a predictive capability based on available use case data, their (hyper) paremeters and technical settings. Currently, development teams face four challenges that complicate the development of ML solutions. First, they lack a formal approach to specify ML solutions that can trace the impact of individual solution components on domain-specific requirements. Second, they lack an approach to document the configurations chosen to build an ML solution, therefore ensuring the reproducibility of the performance obtained. Third, they lack an approach to recommend and select ML solutions that is intuitive for non ML experts. Fourth, they lack a comprehensive sequence of steps that ensures both best practices and the consideration of technical and domain-specific aspects during the development process. Overall, the inability to address these challenges leads to longer development times and higher development costs, as well as less suitable ML solutions that are more difficult to understand and to reuse. This dissertation presents concepts to address these challenges. They are Axiomatic Design for Machine Learning (AD4ML), the ML solution profiling framework and AssistML. AD4ML is a concept for the structured and agile specification of ML solutions. AD4ML establishes clear relationships between domain-specific requirements and concrete software components. AD4ML specifications can thus be validated regarding domain expert requirements before implementation. The ML solution profiling framework employs metadata to document important characteristics of data, technical configurations, and parameter values of software components as well as multiple performance metrics. These metadata constitute the foundations for the reproducibility of ML solutions. AssistML recommends ML solutions for new use cases. AssistML searches among documented ML solutions those that better fulfill the performance preferences of the new use case. The selected solutions are then presented to decision-makers in an intuitive way. Each of these concepts was evaluated and implemented. Combined, these concepts offer development teams a technology-agnostic approach to build ML solutions. The use of these concepts brings multiple benefits, i. e., shorter development times, more efficient development projects, and betterinformed decisions about the development and selection of ML solutions.Item Open Access Flexible and efficient data mapping for simulation of coupled problems(2026) Schneider, David; Uekermann, Benjamin (Jun.-Prof. Dr.)Multi-physics simulations model various physical phenomena and their interactions. Examples include climate models or the simulation of fusion reactors. Modeling more physical phenomena in the same simulation often provides new insights. This poses significant challenges for the underlying methods and the simulation software itself. Decomposing a multi-physics simulation into its parts is an effective way to tame the inherent complexity. The coupling library preCICE allows for such partitioned simulations, coupling simulation models without access to their internal numerics. While preCICE is well-suited for conventional, mesh-based surface couplings, its applicability breaks down in alternative scenarios due to restrictive data-mapping algorithms. These algorithms apply spatial mapping operators to statically defined coupling meshes and are generally not designed for large problems. As a result, they hinder the flexible and efficient simulation of relevant applications, including volumetric couplings, high-order couplings, and mesh-particle couplings. To overcome these limitations, this work focuses on four main aspects: first, developing a scalable partition-of-unity radial-basis-function interpolation customized for coupled problems; second, implementing data-parallel kernel methods on CPUs and GPUs to ensure cross-platform efficiency; third, enabling immediate access to meshes received from coupling partners for user-defined mapping operators; fourth, computing a mapping operator just-in-time on temporary coordinates for the seamless coupling of meshless solvers. Taken together, these concepts enable plug-and-play integration of diverse numerical models in multi-physics simulations. Large-scale volumetric couplings are now feasible and efficient, breaking the traditional accuracy-efficiency trade-off. Multi-physics couplings can exploit spatial high-order convergence rates of existing models for high-fidelity simulations, while preserving full black-box compatibility. The simulation of fluid-particle couplings can be modularized and leverage already-existing models for both the simulation of the mesh-based fluid and the particles. In the end, the robust implementation, the gained efficiency, and the flexibility significantly extend the applicability of preCICE and benefit its vibrant user community. Beyond preCICE, the presented concepts provide generally applicable building blocks for scalable, modular multi-physics coupling.Item Open Access Time-sensitive converged networks : a comprehensive architecture approach(2023) Hellmanns, David; Rothermel, Kurt (Prof. Dr. rer. nat.)Item Open Access B-splines on sparse grids for uncertainty quantification(2021) Rehme, Michael F.; Pflüger, Dirk (Prof. Dr.)Item Open Access Data-efficient and safe learning with Gaussian processes(2020) Schreiter, Jens; Toussaint, Marc (Prof. Dr. rer. nat.)Data-based modeling techniques enjoy increasing popularity in many areas of science and technology where traditional approaches are limited regarding accuracy and efficiency. When employing machine learning methods to generate models of dynamic system, it is necessary to consider two important issues. Firstly, the data-sampling process should induce an informative and representative set of points to enable high generalization accuracy of the learned models. Secondly, the algorithmic part for efficient model building is essential for applicability, usability, and the quality of the learned predictive model. This thesis deals with both of these aspects for supervised learning problems, where the interaction between them is exploited to realize an exact and powerful modeling. After introducing the non-parametric Bayesian modeling approach with Gaussian processes and basics for transient modeling tasks in the next chapter, we dedicate ourselves to extensions of this probabilistic technique to relevant practical requirements in the subsequent chapter. This chapter provides an overview on existing sparse Gaussian process approximations and propose some novel work to increase efficiency and model selection on particularly large training data sets. For example, our sparse modeling approach enables real-time capable prediction performance and efficient learning with low memory requirements. A comprehensive comparison on various real-world problems confirms the proposed contributions and shows a variety of modeling tasks, where approximate Gaussian processes can be successfully applied. Further experiments provide more insight about the whole learning process, and thus a profound understanding of the presented work. In the fourth chapter, we focus on active learning schemes for safe and information-optimal generation of meaningful data sets. In addition to the exploration behavior of the active learner, the safety issue is considered in our work, since interacting with real systems should not result in damages or even completely destroy it. Here we propose a new model-based active learning framework to solve both tasks simultaneously. As basis for the data-sampling process we employ the presented Gaussian process techniques. Furthermore, we distinguish between static and transient experimental design strategies. Both problems are separately considered in this chapter. Nevertheless, the requirements for each active learning problem are the same. This subdivision into a static and transient setting allows a more problem-specific perspective on the two cases, and thus enables the creation of specially adapted active learning algorithms. Our novel approaches are then investigated for different applications, where a favorable trade-off between safety and exploration is always realized. Theoretical results maintain these evaluations and provide respectable knowledge about the derived model-based active learning schemes. For example, an upper bound for the probability of failure of the presented active learning methods is derived under reasonable assumptions. Finally, the thesis concludes with a summary of the investigated machine learning problems and motivate some future research directions.Item Open Access Deep learning aided clinical decision support(2023) Schneider, Rudolf; Staab, Steffen (Prof. Dr.)Medical professionals create vast amounts of clinical texts during patient care. Often, these documents describe medical cases from anamnesis to the final clinical outcome. Automated understanding and selection of relevant medical records pose an opportunity to assist medical doctors in their day-to-day work on a large scale. However, clinical text understanding is challenging, especially when dealing with clinical narratives such as nursing notes or diagnostic reports. These clinical documents differ extensively in length, structure, vocabulary, and lexical and grammatical correctness. In addition, they are highly context-dependent. For all these reasons, approaches based on syntactic rules and discrete text representation often fail to address the variety of clinical narratives propagating unrecoverable errors to downstream applications. Therefore, this thesis focuses on evaluating and designing methods and models that are generalizable and adaptable enough to deal with these challenges. Our goal is to enable text-based clinical decision support systems to utilize the knowledge from clinical archives and medical publications. We aim to design methods that can scale up to the growing amount of clinical documents in hospital archives. A fundamental problem in achieving deep-learning-enabled clinical decision support systems is designing a patient representation that captures all relevant information for automated processing. We engage these challenges by designing a framework for deep-learning-enabled differential diagnosis support. Guided by the needs emerging from this framework, we design and evaluate methods based on three information representation paradigms: (1) Discrete relation extraction using the open information extraction paradigm. (2) Neural text representations based on language and topic modeling. (3) Combining complementary neural text representations. Our framework translates clinical diagnostic steps and pathways to statistical and deep-learning-based models. Accordingly, we can show that deep-learning-enabled differential diagnosis benefits from contextualized information representations. Further, we identify shortcomings of the open information extraction paradigm in a comprehensive benchmark. We design a distributed text representation model based on topical information. Our extensive large-scale experiment results show that topical distributed text representations capture information complementary to language modeling-based approaches across domains, thus enabling a holistic text representation for medical texts. Our experiments with medical doctors using our prototypical implementation of the deep-learning-enabled differential diagnosis process validate this framework. Moreover, we identify seven crucial design challenges for text-based clinical decision support systems based on our qualitative and quantitative findings.Item Open Access Data engineering concepts, framework, and algorithms for personalized and transparent decision support(2025) Oppold, Sarah; Herschel, Melanie (Prof. Dr.)Machine learning models are commonly used for decision support even though they are far from perfect, e.g., due to bias introduced by imperfect training data or wrong feature selection. While efforts are made and should continue to be put into developing better models, we will likely continue to rely on imperfect models in many applications. In this thesis, we follow a rationale similar to the practices that emerged in medicine to develop decision support systems (DSS) that are demonstrably developed as responsibly as possible. Therefore, we present novel approaches to develop personalized and transparent data-driven decision support systems. Our fist contribution is a novel holistic system framework that offers transparent and personalized services tailored to user profiles to serve their best interest. Our framework personalizes the choice of model for individuals or groups of users based on metadata about datasets and machine learning models. Querying and processing these metadata ensures transparency by supporting various kinds of queries by different stakeholders. We discuss our framework in detail, show why existing solutions are inadequate, and highlight research questions that need to be addressed in the future. Based on a prototypical implementation, we showcase that even a baseline implementation of our framework supports the desired transparency and personalization. To address the problem of personalization, we present novel algorithms for dynamic, fair, and accurate decision support systems in our second contribution. We first propose a general procedure that combines context-sensitive algorithms for generating dynamic model ensembles necessary for personalization with (static) fair and accurate model ensembles, satisfying our multi-objective goal. We further introduce a family of algorithms, jointly called FALCES, representing different alternatives for our procedure. Using the FALCES algorithms, we evaluate our framework for dynamic, fair, and accurate model ensembles on synthetic and real data. The results show that despite the presence of biases, our algorithms outperform state-of-the-art fairness algorithms while maintaining acceptable accuracy. In order to provide personalized and transparent decision support, we have to capture and manage a large and diverse set of metadata on datasets, which are the foundation of decision support systems. Therefore, we present LiQuID in our third contribution. LiQuID is the first systematic and holistic metadata model for accountable datasets, i.e., datasets on which queries of an accountability workload can be answered. We evaluate LiQuID by comparing it to existing metadata standards from the responsible data analysis community and more matured disciplines. In a second evaluation, we compare LiQuID to a new workload for accountable datasets that we created based on a vast survey of the GDPR, an FTC report, and dataset expert interviews. Finally, we look at the devised system’s usefulness in achieving the desired goals. In our fourth contribution, we delve into one particular setting that has been claimed by research before, where transparency potentially serves to improve trust. We critically review the term “trust” to define a theoretical model for trust in data engineering. Based on this model, we describe a framework for trust in data engineering that integrates trust in the data engineering pipeline and serves as a guideline to develop a trust strategy. We also describe a general procedure to evaluate the effects of a transparency measure on trust. Finally, we apply and evaluate our methods on a credit scoring use case. Results show that transparency not necessarily increases trust, highlighting the importance of a more systematic study of the problem using our proposed methods. The research presented in this thesis has significantly contributed to our goal for decision support systems that are demonstrably developed as responsible as possible. This includes a general system framework for personalized and transparent decision support, novel optimization algorithms, accountability metadata models and a critical discussion on using transparency to foster trust.Item Open Access Computational methods for partitioned simulation coupling : applications in multi-physics simulations and energy infrastructure optimisation(2023) Davis, Kyle; Schulte, Miriam (Prof. Dr.)