05 Fakultät Informatik, Elektrotechnik und Informationstechnik
Permanent URI for this collectionhttps://elib.uni-stuttgart.de/handle/11682/6
Browse
Search Results
Item Open Access Speech interface for human and robot collaboration(2018) Kashif, Moin UddinIn the past, robots and machines were mostly designed to perform specific tasks without much human interaction needed. Nowadays with the advancements in technology, intelligent robots can be designed which can perform multiple tasks, interact with the surrounding environment, assist and give valuable suggestions to humans etc. so an efficient and natural mode of communication is required for this human-robot interaction. In this thesis, we proposed an architecture to develop a speech interface for human-robot interaction. The speech interface is used to give voice commands to the robot, PR2, in order to perform 5 tasks which are designed to test the performance of the speech interface. The tasks are sorting, shaping, stacking, building and balancing of 6 objects on table-top which are designed and ordered by the level of difficulty. First two tasks are comparatively easier as the user doesn't have to follow any order to finish them, next two tasks require to follow the order and in the last task, the stack of objects must be balanced in order to finish it. The speech interface receives voice commands from the user, convert them into text, maps to the corresponding command and send to the task manager to perform the operation. After that, it processes the received command, takes the appropriate decision based on the current status of the task and available actions and sends the command to the PR2 to perform the operation. Additionally, we have designed a feedback mechanism where PR2 sends back the feedback to the task manager which is delivered back to the speech manager so that it can be converted into an audio signal and play for the user. Furthermore, the system uses a TCP connection for the exchange of data and information between the speech manager and the task manager. The speech interface is also compared with other modalities such as text input and graphical user interface with the same tasks and we have also conducted user study to evaluate the system performance. The results show that the participants prefer speech interface as it feels more natural.Item Open Access Orthogonale Dünngitter-Teilraumzerlegungen(2018) Schreiber, ConstantinIn der Simulation treten Häufg hochdimensionale partielle Differentialgleichungen auf. Das Lösen dieser wird für volle Gitter sehr schnell zu teuer. In dieser Arbeit wird ein Verfahren für das Lösen partieller Differentialgleichungen mit Hilfe von Dünnen Gittern, welche für mehrdimensionale Probleme besser skalieren, sowie dessen Implementierung in das Programmpaket SG++ vorgestellt. Durch Funktionsdarstellung in einem Erzeugendensystem wird die Verwendung einer L2-orthogonalen Teilraumzerlegung ermöglicht. Projektionsoperatoren ersetzen hierbei die explizite Transformation in eine Prewavelet-Basis. Diese Zerlegung erlaubt das Lumping der Steifgkeitsmatrix, also das Weglassen von großen Blöcken der Matrix. Hiermit wird ein Algorithmus zur Matrixmultiplikation, welcher dem von Schwab und Todor ähnelt implementiert. Dieser wird in einem konjugierten Gradienten-Verfahren verwendet und auch auf krummberandete Gebieten angewendet. Des Weiteren wird die Teilraumzerlegung durch L2-Projektion mit anderen Zerlegungen in Bezug auf Laufzeit und Fehlerentwicklung verglichen.Item Open Access Vision assisted biasing for robot manipulation planning(2018) Puang, En YenSampling efficiency has been one of the major bottlenecks of sampling-based motion planner. Although being more reliable in complex environments, Rapidly-exploring Random Tree for example often requires longer planning time than its optimisation-based counterpart. Recent developments have introduced numerous methods to bias sampling in configuration-space. Gaussian mixture model, in particular, was proposed to estimate feasible regions in configuration-space for low-variance task. Unfortunately this method does not adapt its biases according to individual planning scene during inference. Therefore, this work proposes vision assisted biasing to adapt biases by changing the weights of Gaussian components upon query. It uses autoencoder to extract features directly from depth image, and the resulted latent code is then used for either nearest neighbours search or direct weights prediction. With a modified pipeline, these extensions show improvements on not only the sampling efficiency but also path optimality of simple motion planner.Item Open Access Modeling recommendations for pattern-based mashup plans(2018) Das, SomeshData mashups are modeled as pipelines. The pipelines are basically a chain of data processing steps in order to integrate data from different data sources into a single one. These processing steps include data operations, such as join, filter, extraction, integration or alteration. To create and execute data mashups, modelers need to have technical knowledge in order to understand these data operations. In order to solve this issue, an extended data mashup approach was created - FlexMash developed at the University of Stuttgart - which allows users to define data mashups without technical knowledge about any execution details. Consquently, modelers with no or limited technical knowledge can design their own domain-specific mashup based on their use case scenarios. However, designing data mashups graphically is still difficult for non-IT users. When users design a model graphically, it is hard to understand which patterns or nodes should be modeled and connected in the data flow graph. In order to cope with this issue, this master thesis aims to provide users modeling recommendations during modeling time. At each modeling step, user can query for recommendations. The recommendations are generated by analyzing the existing models. To generate the recommendations from existing models, association rule mining algorithms are used in this thesis. If users accept a recommendation, the recommended node is automatically added to the partial model and connected with the node for which recommendations were given.Item Open Access Automatic indoor modelling using crowd-sensed point clouds(2018) Prabha Sekar, SuriyaThere are many well-established approaches to model outdoor environments but it has been a challenge to come up with an accurate and reliable approach for indoor modelling. Outdoor automatic modelling and mapping are achieved using satellite positioning systems like GPS. However, for indoor modelling, due to localization problems, positioning systems do not help. Therefore, the indoor positioning is done by combining inertial sensors along with area learning. Previously, the 3D models for indoor environments were done either manually which is a time-consuming process or by using range images obtained from laser range scanners which are an expensive approach. For overcoming this drawback, we suggest utilizing crowd sensing in order to obtain the environment’s spatial information. Crowd-sensing is performed by a group of participants using their mobile devices to execute certain assigned tasks. Nowadays modern mobile devices are used for many crowd sensing applications with the help of the ubiquitous presence of such powerful devices. In our approach, the crowd sensing task involves collecting scans of rooms in public buildings using these mobile devices. Thus, we depend on the new wave of powerful devices, e.g. Google Tango, Microsoft Hololens and Apple ARKit which generates optimal scanned data i.e. 3D point clouds which can be crowd sensed and then processed to automatically generate indoor models. These devices provide capabilities like depth perception, area learning and motion tracking, which help to acquire the spatial information of the room and its relative position. Even though there is more energy consumption on the mobile device which is responsible for collecting large dataset, this can be reduced by applying Octree compression which reduces the amount of data of the scanned point clouds. The scanned data is sent to the server where it gets processed and generates the required indoor model. The proposed tool will derive the 3D indoor model of a floor. The tool is capable of extracting all the planes from the point clouds, detecting all the room surfaces (i.e. at least four walls, a ceiling and a floor), classifying wall openings and classifying high-level semantics (i.e. furniture). The basic model for all the rooms in the floor is generated and in order to build a floor model, the ground truth of that floor is utilized. Finally, the accuracy of the basic model is enhanced using the grammar model fitting tool. Using the point cloud scans from one of our University campuses, the final indoor modelling tool was able to derive the complete 3D floor plan in compliance with the ground truth.Item Open Access A massively parallel combination technique for the solution of high-dimensional PDEs(2018) Heene, Mario; Pflüger, Dirk (Jun.-Prof. Dr.)The solution of high-dimensional problems, especially high-dimensional partial differential equations (PDEs) that require the joint discretization of more than the usual three spatial dimensions and time, is one of the grand challenges in high performance computing (HPC). Due to the exponential growth of the number of unknowns - the so-called curse of dimensionality, it is in many cases not feasible to resolve the simulation domain as fine as required by the physical problem. Although the upcoming generation of exascale HPC systems theoretically provides the computational power to handle simulations that are out of reach today, it is expected that this is only achievable with new numerical algorithms that are able to efficiently exploit the massive parallelism of these systems. The sparse grid combination technique is a numerical scheme where the problem (e.g., a high-dimensional PDE) is solved on different coarse and anisotropic computational grids (so-called component grids), which are then combined to approximate the solution with a much higher target resolution than any of the individual component grids. This way, the total number of unknowns being computed is drastically reduced compared to the case when the problem is directly solved on a regular grid with the target resolution. Thus, the curse of dimensionality is mitigated. The combination technique is a promising approach to solve high-dimensional problems on future exascale systems. It offers two levels of parallelism: the component grids can be computed in parallel, independently and asynchronously of each other; and the computation of each component grid can be parallelized as well. This reduces the demand for global communication and synchronization, which is expected to be one of the limiting factors for classical discretization techniques to achieve scalability on exascale systems. Furthermore, the combination technique enables novel approaches to deal with the increasing fault rates expected from these systems. With the fault-tolerant combination technique it is possible to recover from failures without time-consuming checkpoint-restart mechanisms. In this work, new algorithms and data structures are presented that enable a massively parallel and fault-tolerant combination technique for time-dependent PDEs on large-scale HPC systems. The scalability of these algorithms is demonstrated on up to 180225 processor cores on the supercomputer Hazel Hen. Furthermore, the parallel combination technique is applied to gyrokinetic simulations in GENE, a software for the simulation of plasma microturbulence in fusion devices.Item Open Access Learning to plan in large domains with deep neural networks(2018) Shou, ZhenkaiIn the domain of artificial intelligence, effective and efficient planning is one key factor to developing an adaptive agent which can solve tasks in complex environments. However, traditional planning algorithms only work properly in small domains. Learning to plan, which requires an agent to apply the knowledge learned from past experience to planning, can scale planning to large domains. Recent advances in deep learning widen the access to better learning techniques. Combining traditional planning algorithms with modern learning techniques in a proper way enables an agent to extract useful knowledge and thus show good performance in large domains. This thesis aims to explore learning to plan in large domains with deep neural networks. The main contributions of this thesis include: (1) a literature survey on learning to plan; (2) proposing a new network architecture that learns from planning, combining this network with a planner, implementing and testing this idea in the game Othello.Item Open Access Window drop load shedding in complex event processing(2018) Flaig, AlbertIn Complex Event Processing (CEP) huge input event streams are processed to interpret specific situations in real time. In the field of parallel CEP the event stream is split into different windows which can be processed in parallel by several operator instances. This paradigm reduces the load imposed on a single operator instance and allows for horizontal scalability. However, in case of high load, the operator instances may not be able to process the incoming events in time; the unprocessed events are queued up resulting in a higher processing latency. In such situations it can be desirable to impose a latency bound on the system. One way to satisfy the given latency bound, is to do load shedding by dropping incoming windows. This problem is not trivial, as it is unclear when, which and how many windows need to be dropped. To answer these questions we try to find out the most promising windows, e.g. windows which may yield a high amount of complex events, yet impose a low impact on processing time. However, this adds additional challenges, as the processing latency of a window depends on many unknown variables, such as the size of the window, the types of incoming events, the position of these events relative to each other and even on the processing latency of other overlapping windows. Furthermore, even if the quality and processing latency of a window is determined, there are several open questions regarding the timing and frequency of load shedding in order to not violate the latency bound, but still keep enough windows active. In the scope of this thesis, we introduce a latency and quality model to estimate the processing latency a window induces. Based on this model, we propose an algorithm which decides the windows to drop in case of high system load to satisfy a given latency bound while minimizing the loss of quality.Item Open Access Stückweise lineare Funktionen für Network Calculus(2018) von Zameck Glyscinski, PeterNetwork Calculus ist ein häufig verwendetes und mächtiges Tool, um Aussagen über die Performanz und mögliche Probleme in Kommunikationsnetzwerken treffen zu können. Mit Hilfe dieser Aussagen zur Performanz eines Kommunikationsnetzwerkes ist es möglich, vorab Anforderungen an ein Kommunikationsnetzwerk zu stellen und dessen Aufbau richtig zu planen. Es ist dadurch auch möglich die maximale Belastung eines bestehenden Kommunikationsnetzwerkes zu bestimmen, ohne es einem tatsächlichen Stresstest unterziehen zu müssen. Das mathematische Grundgerüst des Network Calculus basiert dabei auf der (min,+) Algebra. Um die einzelnen Operationen der (min,+) Algebra berechnen zu können, bedarf es einer Darstellung der einzelnen Ankunftskurven und Servicekurven in dem Kommunikationsnetzwerk. Eine mögliche Art diese Ankunftskurven und Servicekurven darzustellen, sind stückweise lineare Funktionen. In dieser Arbeit wird basierend auf der Arbeit von Bouillard und Thierry eine Pythonbibliothek für die effiziente Berechnung mit stückweisen linearen Funktionen in der (min,+) Algebra erstellt. Die richtige Datenstruktur zum Darstellen einer stückweisen linearen Funktion, sowie Algorithmen zum Berechnen der Operationen in der (min,+) Algebra, werden erläutert. Wichtige Berechnungen wie die Faltung und Entfaltung, das Minimum und Maximum und die Addition und Subtraktion, sowie das Backlog und das Virtual Delay von zwei stückweisen linearen Funktionen werden dabei unterstützt.Item Open Access Evaluation of prediction mechanisms of parameters for data mining algorithms(2018) Loutfi, KindaExtracting knowledge and useful information from huge amount of data is one of the biggest issues currently in the world of computer science. Data mining is one essential step in the knowledge discovery process due to the importance of its contribution in extracting this knowledge. One of the most famous mining techniques is Clustering, which is a widely used approach in data mining. Clustering is the process of partitioning a group of objects into smaller sets called clusters, in which similar objects are assigned to the same cluster, and dissimilar objects are assigned to different clusters. K-Means, DBSCAN, and OPTICS are three of the most popular clustering algorithms which are used to group similar data into clusters. Each of these algorithms requires input parameters. The difficulty of knowing these input parameters in advance is the flaw of these algorithms. Many previous approaches were introduced which provide a prediction of these parameters. However, different problems emerged while using these approaches, such as a long overall runtime to achieve predicted parameter values, since the clustering algorithm is applied multiple times with varying parameters to identify the best parameter configuration. This thesis introduces a new approach for predicting the input parameters. The proposed approach is called PROD (Position-Based Prediction Using Voronoi Diagrams). PROD overcomes the problems which emerged in previous approaches. The prediction of the input parameters is performed by this new approach through using space partitioning approaches and subsequently exploiting their properties. PROD is evaluated on various data sets. Regarding K-Means algorithm; PROD provides a prediction of number of clusters as well as a prediction of the initial centroids. The experimental results unveil, that PROD is (a) more accurate and (b) faster by a factor of 26.5 in contrast to previously best existing approaches. Additionally, the results of the evaluations show that it can be used either as a prediction approach for the amount of clusters or as a standalone clustering algorithm. Despite the effectiveness of the prediction made by PROD with regard to the K-Means parameters, the results of the experiments show that this novel approach needs further improvement to make the prediction of DBSCAN and OPTICS parameters work. More clearly, changing some aspects of PROD might lead to a better prediction with regard to these two algorithms.