Bidirectional RNN is now available in Tensorflow

Machine learning




Bachelor thesis by Mulham Alesali

Conception and implementation of a neuroevolutionary algorithm for controlling a vehicle in Unity

Genetic algorithms (GA) can be used to optimize the weightings of artificial neural networks (kNN). The learning task in this case is to find a policy that is able to control a vehicle in a simple simulated environment and thus belongs to reinforcement learning. For this purpose, the GA is to be implemented and applied to the learning task. The (physics-based) simulation is to be developed in a suitable manner in Unity and should take into account aspects of conveying concepts. This would be conceivable through the visualization of the genotypes, the fitness distribution or the fitness development. The particular difficulty lies in the development of an overall application with simulation, process control, visualization and AI components.

The result was a clear, motivating Unity / C # application that evolves a population of neural networks as a policy for controlling vehicles. The evolution can be followed over the generations by means of fitness distribution and fitness curves. The relationships are made clear by color codes in the genotype. Trained populations can then be transferred to other routes and perform similarly there.

Colloquium: March 29, 2021

Supervisor: Dipl.-Inform. Ingo Boersch, Prof. Dr. Jochen Heinsohn

Download: A1 poster, thesis

Bachelor thesis by Mahmoud Abdelrahman

Benchmarking Post-Training Quantization for Optimizing Machine Learning Inference on compute-limited Edge Devices

Edge AI, i.e. the transfer of intelligence from the cloud to edge devices such as smartphones and embedded systems, has gained in importance in recent years. This requires optimized machine learning (ML) models that can work on computers with limited computing power. Quantization is one of the essential techniques of this optimization. The data type for displaying the parameters of a model is changed here. In this thesis, quantization was examined, in particular the post-training quantization techniques available in TensorFlow Lite (TFLite). An image classification model trained on the MNIST data set and a semantic segmentation model trained on the Cityscapes data set were used to carry out experiments. For the benchmarking, the inference was carried out on two hardware-different CPU architectures, namely on a laptop and a Raspberry Pi. For the benchmarking, metrics such as model size, accuracy, mean intersection over union (mIOU) and inference speed were handled. For both image classification and semantic segmentation models, the results showed an expected reduction in model size when different quantization techniques were used. In both cases, the accuracy and mIOU have not changed significantly from that of the original model. In some cases, the use of quantization even improved the accuracy. The speed of inference with regard to the image classification model has improved adequately. In some cases, the inference speed on Raspberry Pi even increased by a factor of 10.

Colloquium: 02.03.2021

Supervisor: Prof. Dr.-Ing. Jochen Heinsohn, Abhishek Saurabh (MSc) Volkswagen Car.Software Organization, Dipl. Inform. Ingo Boersch

Download: A1 poster, bachelor thesis

Bachelor thesis by Bhirawa Satrio Nugroho

Thursday, February 11, 2021

Performance optimization in machine learning using the example of credit checks for bank customers

The credit check is an important step performed by lenders that can determine whether or not the banking institution will provide credit to potential borrowers. This exam has a huge impact on agencies, especially in the financial sector. To avoid financial problems that arise due to lending risks, what is needed is a method that aids in credit scoring by increasing the statistical performance of a credit scoring model. With the help of machine learning models, the time, effort and cost of performing statistical analyzes that are applied to big data can be reduced. For this reason, machine learning algorithms, namely from Logistic Regression, K-Nearest Neighbors and Support Vector Machine, are compared in this thesis. Experiments are also conducted that can improve the performance of these models.

Colloquium: 02/11/2021

Supervisor: Prof. Dr. Jochen Heinsohn, Dipl.-Inform. Ingo Boersch

Download: A1 poster

Bachelor thesis by Sebastian Tillack

Thursday, February 11, 2021

Decision support with Bayesian networks - modeling of a COVID-19 domain with HUGIN

Bayesian networks (BN) are well suited for modeling uncertainty. A current example of the occurrence of uncertainty is the COVID-19 domain, in particular the relationships between symptoms, analyzes, effects and consequences. After a brief introduction to the basics of BN, the essential concepts of the COVID-19 domain including their interrelationships will be presented. This is followed by an analysis of the state of research on BN, which already has this domain as an application, and also a short assessment of its own. The core of the bachelor thesis is its own implementation with the help of the HUGIN tool.

Results

The resulting application makes it possible to determine the likelihood of illness from COVID-19, SARS, MERS or influenza. For this, the observed symptoms are made known to the network as evidence. This means that the value of the corresponding variable is fixed and is no longer dependent on the original probability. It can be shown that specific symptoms, such as impaired taste and / or smell, influence the posterior probabilities of the disease more than common symptoms such as cough.

Colloquium: 02/11/2021

Supervisor: Prof. Dr. Jochen Heinsohn, Dipl.-Inform. Ingo Boersch

Download: A1 poster

Bachelor thesis by Robert Beilich

Friday, October 23, 2020

Tooling for big data extraction

This thesis presents problems and solutions that can occur when working with large unstructured data sets. This is done using the practical example of extracting the JavaScript libraries that have been used over time from the CommonCrawl data set. Starting with few hardware resources and later using the stronger infrastructure of the Future SOC Lab, the various problems that these development stages bring with them are dealt with, for example scarce resources for operating the database and the hardware configuration. Finally, the collected findings are implemented using part of the data set for the practical example and the results are visualized. The restriction to only part of the data record results from the fact that the entire data record cannot be processed with the existing hardware.

Colloquium: October 23, 2020

Supervisor: Prof. Dr. Sven Buchholz, Dipl. Inform. Ingo Boersch

Download: A1 poster

Master project AI: islands of evolution

Monday, September 28, 2020

Evolution Islands - Part 1

This sub-project is about the simulation of evolution in an artificial ecological system. The long-term goal is a clear visualization of various evolutionary phenomena such as gene drift, multi-criteria optimization, dynamic fitness landscapes, coevolution, convergent evolution and others. These effects are difficult to observe in nature because of their distribution and slowness and should be experienced by individuals in a 3D world. It should be fun to follow the experiments like an exciting film. The project is based on the 2D world in [1].

Phase 1: Closing the evolutionary loop in a simple ecology

Result

Executable prototype with genotypes, phenotypes, movement, mutation, inheritance and food (UnrealEngine, C ++ and Blueprint)

Source text and exe on request to [email protected]

[1] Ventrella J. (2005) Gene Pool: Exploring the Interaction Between Natural Selection and Sexual Selection. In: Adamatzky A., Komosinski M. (eds) Artificial Life Models in Software. Springer, London. https://doi.org/10.1007/1-84628-214-4_4

Bachelor thesis by Rick Lüdicke

Thursday, July 09, 2020

Exploratory analysis and data-based modeling of a forecast model to determine the monthly cost burden in the RentSharing model

The aim of the work is to take the first steps to expand an offer platform for leasing contracts with a suggestion system. The previous system uses a so-called company car computer (PWR) to calculate the monthly cost for the contract data to be entered by the user. The calculation is time consuming. By accelerating the PWR, the target size could already be calculated with partially entered contract data for a multitude of options, for example vehicle types, and thus serve as the basis for a proposal for a contract feature. This thesis tries to achieve the goal through a data-based approximation of the PWR. The main goals of the work are thus:

  • Create the dataset
  • Definition of the target size
  • Exploratory analysis
  • Exploratory analysis related to the target variable
  • optional: determination of relevant characteristics, selection of data records, preparation of characteristics, definition of new characteristics
  • Concept of the evaluation process
  • Model creation and evaluation
  • Discussion of the models

One difficulty of the task is embedding it in a real company context, as well as in the special situation of the COVID pandemic.

Colloquium: 07/09/2020

Supervisor: Dipl. Inform. Ingo Boersch, Prof. Dr. Susanne Busse

Download: A1 poster

Master thesis by Darya Martyniuk

Wednesday, January 22, 2020

Combination of imitation learning and reinforcement learning for movement control

A successful combination of imitation learning (IL) and reinforcement learning (RL) to control the motion of a robot has the potential to provide an end user without programming knowledge with an intelligent robot that is able to learn the required motor skills from humans and to adapt them independently in view of the current framework conditions and goals. In this master's thesis, a combination of IL and RL is used to control the movement of the humanoid robot NAO. The learning process takes place on the real robot without prior training in a simulation. Laying the foundation for learning kinesthetic demonstrations by an expert as well as the agent's own experience gained from interacting with the environment.

The learning process used is based on the algorithms Deep Deterministic Policy Gradient from Demonstration (DDPGfD) and Twin Delayed Policy Gradient (TD3) and is evaluated in a case study, the game Ball-in-a-Cup. The results show that the implemented algorithm enables efficient learning. Pre-trained with the data from demonstrations, the robot begins interacting with the environment with a suboptimal strategy, which it quickly improves in the course of the training. However, the performance of the algorithm is highly dependent on the con fi guration of the hyperparameters. In future work, a simulation is to be created for the ball-in-a-cup game, in which the hyperparameters and possible improvements in the learning process can be evaluated before training with the real robot.

Video: Execution of the optimal policy after 200 learning episodes

Colloquium: January 22, 2020

Reviewer: Dipl.-Inform. Ingo Boersch, Prof. Dr.-Ing. Jochen Heinsohn

Download: A1 poster, master's thesis

Master thesis by Mario Kaulmann

Tuesday, October 29, 2019

Application of interactive evolutionary algorithms to generate drum rhythms

Finding interesting drum rhythms is a creatively demanding task. There are options to play different instruments at different times. The arrangement of the instruments to be played over time must be repeatable and please the drummer. Interactive evolutionary algorithms are proposed to support this process. The interactivity allows the user to control the search and the modification operators of the evolutionary algorithm generate new suggestions. One focus of this work is the user interface. This should keep user fatigue low and also contribute to the traceability of the evolutionary algorithm. Theoretical fundamentals are explained, inspiring work is considered, the conception and implementation of a demonstration program is described and experiments with the program are documented and evaluated.

Evolutionary algorithms are optimization processes that are inspired by the evolution of living things. Solution proposals (individuals) are generated, which are assigned an evaluation (fitness), on the basis of which a selection (selection) takes place which individuals are taken to generate the next individuals. The coding of an individual is the genotype, the appearance of the individual is the phenotype. If there is a human-machine interface, then it is one interactive evolutionary algorithm according to the extended definition from [Tak01]. The inclusion of people in the process is the weak point of this approach, as people tire quickly from constant activities. In addition, the comparison of several time-sequential individuals represents a particular cognitive burden on the user.

[Tak01] Takagi, H .: Interactive evolutionary computation: fusion of the capabilities of EC optimization and human evaluation. In: Proceedings of the IEEE 89 (2001), No. 9. http://dx.doi.org/10.1109/5.949485. - DOI 10.1109 / 5.949485

Colloquium: October 29, 2019

Reviewer: Dipl.-Inform. Ingo Boersch, Prof. Dr. Martin Christof Kindsmüller

Download: A1 poster, master's thesis

Bachelor thesis by Katharina Geue

Wednesday, September 18, 2019

Usage-based optimization of motorcycle tours with map matching technologies

In this work, anonymized, recorded motorcycle tours (tracks) by calimoto users are to be analyzed in order to determine how often motorcyclists have ridden on which roads. A new routing profile should be created from this, which should generate routes over the most popular roads. It is also evaluated whether the integration of the frequency values ​​in the route planning is useful and whether this generates routes suitable for motorcyclists.

Colloquium: September 18, 2019

Supervisor: Prof. Dr.-Ing. Jochen Heinsohn, Sebastian Dambeck M.Sc. (calimoto GmbH), Dipl.-Inform. Ingo Boersch

Download: A1 poster

Bachelor thesis by Hüsein Celik

Thursday, September 05, 2019

Re-implementation of a U network for segmentation with the GluonCV framework

The thesis examines the performance and applicability of GluonCV (frameworks for the simplified use of deep neural networks) for the semantic segmentation of medical image data with a U-Net. For this purpose, based on the U-Net architecture [RFB15], an existing Keras implementation [Pet18] is reimplemented in GluonCV and various criteria for feasibility and performance are evaluated in systematic experiments.

[RFB15] Ronneberger, Olaf; Fischer, Philipp; Brox, Thomas: U-Net: Convolutional Networks for Biomedical Image Segmentation. In: CoRR (2015).

[Pet18] Petsiuk, Vitali: Lung segmentation (2D). (December 2018). https://github.com/imlab-uiip/lung-segmentation-2d, accessed: July 5, 2019.

Colloquium: 09/05/2019

Reviewer: Prof. Dr. Sven Buchholz, Dipl.-Inform. Ingo Boersch

Film: Darya on her master's project

Short film "Intelligent Systems": Darya on her master's project II

Darya explains her master's project in the 2nd semester on the subject of "Reinforcement learning with opponents" with a focus on intelligent systems
The aim of project II is the further development of the application started in project I, which enables a NAO robot to play the NIM game with a human opponent. The robot should independently recognize the game situation, plan the strategy, make its own moves and interact with the opponent.

The following tasks are to be solved in project II:
  • Image processing: Automatic recognition of the region of interest.
    In project I, the coordinates of the region of interest (abbr .: ROI) are fixed. As part of project II, the ROI should be determined automatically
  • Image processing: Recognition of the playing field condition.
    There is no predetermined order in which the Lego bricks should be pushed away. This is why the robot should recognize exactly which of the Lego bricks are still in the game and which are not.
  • Image processing: determining the end of the game.
    The method implemented in Project I to determine the end of the game by the opponent was based on hand recognition in the ROI. However, there are several problems with this approach. For this reason, instead of the currently used method, the idea of ​​observing the number of Lego bricks in the ROI to determine the end of the game by the user should be implemented. The robot should be able to detect fraud and, if necessary, issue a warning to the opponent.
  • Interaction: design of a human-like behavior of the robot.
    The robot should show lively and committed behavior. This means that the robot should perceive the user, have a dialogue with him, gesticulate and show the emotions.
  • Strategy: The independent learning of a game strategy.
    The robot must try to win the game. With the help of a method from the family of reinforcement learning algorithms, the robot must determine an optimal strategy and follow it or adapt to a new opponent.

Master thesis by Eric Bunde

Sentiment analysis using German Twitter corpora and deep learning

The aim of this work is to investigate approaches for analyzing the sentiment of tweets in German. For this purpose, a data mining process has to be run through with the phases of data selection, exploration, preparation, feature generation, modeling and evaluation. The particular difficulty lies in the data acquisition, the unstructured data (tweets) as well as the selection and executable implementation of the learning algorithms from the area of ​​deep learning. The classifiers and their hyperparameters should be documented in a comprehensible manner and suitably evaluated.

Colloquium: March 29, 2019

Supervisor: Prof. Dr. Sven Buchholz, Dipl.-Inform. Ingo Boersch

Download: A1 poster, master's thesis

Master thesis by Joel Rixen

Friday September 28, 2018

Automatic generation of fingering for piano scores using machine learning

When learning to play the piano, finding a good fingering can be problematic, especially for newcomers. The aim of this work is to create a program that can generate the fingering for piano scores.

Attempts have been made to solve this problem several times over the past 20 years. The solution approaches are mostly based on the same idea and only work for simple piano pieces. For this reason, a different approach (machine learning) was used in this thesis. An application for generating fingering from scores with the current method of the bidirectional LSTM networks was designed and successfully implemented.

The particular difficulty lay in the complexity of the chosen application scenario and the complex tests to evaluate the performance. Complex because there can be several easily playable fingerings for a score, so that the automatically generated fingerings have to be evaluated by actually playing.

Colloquium: 09/28/2018

Supervisor: Prof. Dr. rer. nat. Martin Christof Kindsmüller, Dipl.-Inform. Ingo Boersch

Download: A1 poster

Master thesis by Herval Bernice Nganya Nana

Monday, September 24, 2018

Multi-staged Deep Learning approach for automatic counting and detecting banana trees in UAV images using Convolutional Neural Networks

The thesis addresses a difficult problem in the automated monitoring of plant conditions on farms and plantations. She proposes a method based on deep learning to automatically detect and count banana trees on a banana plantation using drone images. The difficulty with this task is that banana tree tops overlap very often. Even for a human, this task is very difficult to do. The task is thus divided into two parts: localization (detection) and counting.

A multi-stage approach is used to solve this problem: a classifier detects whether an ROI contains a banana tree crown, a following regressor determines the coordinates of the crown candidates in the ROI. A final aggregation summarizes the tree canopy candidates into recognized tree centers. In the end, these are the tree tops you are looking for. The developed method is evaluated on a test case with promising results. In this test case, where there are densely planted trees, the model achieved a margin-of-error of 0.0821. This corresponds to a quality of 91.79% for the counting task. The low average distance error of around 43 cm for the localization task should be emphasized.

Colloquium: 09/24/2018

Supervisor: Prof. Dr.-Ing. Sven Buchholz, Dipl.-Inform. Ingo Boersch, Jan Vogt (Orca Geo Services GmbH, Brandenburg)

Download: A1 poster

Bachelor thesis by Ursina Bisang

Online deep learning with hedge backpropagation for predictive maintenance applications

The aim of the thesis is to investigate the suitability of hedge backpropagation for predicting machine failures on the turbofan data set. Hedge backpropagation is a multilayer perceptron in which an additional output is generated from each hidden layer. The outputs are weighted linearly and their weighting is adapted to the rest of the network using the hedge algorithm. So it should be possible that the used depth of the network automatically adapts to the task via these weightings.

The approach should be presented in detail and implemented either by itself or with the help of a suitably chosen implementation and compared with other approaches, e.g. LSTM networks, according to sensibly chosen criteria.

Colloquium: 08/06/2018

Supervisor: Dipl.-Inform. Ingo Boersch, Prof. Dr.-Ing. Sven Buchholz, Prof. Dr. rer. nat. Adrian Paschke (Fraunhofer FOKUS)

Download: A1 poster

Master thesis by Jonas Preckwinkel

Deep learning for object detection in images with region-based convolutional neural networks and GPU computing

The aim of the work is to investigate the performance of current artificial neural networks for object detection in images. For this purpose, essential concepts of deep learning are to be explained in the introduction. The focus of the work is the evaluation of region-based object detection systems. For this purpose, the functionality of multi-level, region-based detection systems, in particular R-CNN, Fast-RCNN and Faster-RCNN, will be explained and compared in detail. The networks are to be implemented and trained on suitable data, e.g. the VOC data volumes, and the influence of hyperparameters on computing times and performance must be examined. The evaluation scenarios and performance criteria are to be selected appropriately.

Colloquium: 07/11/2018

Supervisor: Dipl.-Inform. Ingo Boersch, Prof. Dr. Sven Buchholz

Download: A1 poster, master's thesis

Master thesis by Vanessa Vogel

Human-Robot-Interaction for the supervised learning of object recognition by the humanoid robot NAO

The aim of the work is to develop an application to demonstrate the interaction of a person with a robot to learn about objects that are to be recognized later. The focus is on the one hand on the development of the interaction model and on the other hand on reliable recognition after a few learning examples. The particular difficulty lies in the use of the NAO robot. The work should take into account the preparatory work and develop it further.

The result was a Python application that is able to label and recognize objects in dialogue with a person. A creative solution is the robust input of the label via stamp and character recognition. Recognition is achieved through segmentation and classification. The segmentation takes place pragmatically on the basis of an initial background, for the classification SIFT features are extracted as object features and thus an RBF-SVM is trained.

Colloquium: 07.03.2018

Supervisor: Dipl.-Inform. Ingo Boersch, Prof. Dr. Jochen Heinsohn

Download: A1 poster

Master thesis by Colin Christ

Friday February 02, 2018

Real-time reinforcement learning of action strategies for humanoid robots

The aim of the work is to develop an application to demonstrate reinforcement learning (RL) on autonomous, humanoid robots. The aim is to demonstrate the learning of a successful action strategy in a simple real scenario. You can choose the scenario yourself, e.g. sorting balls. The scenario should essentially be deterministic, but in rare cases it can react stochastically. The learning process should be able to run independently and unattended and lead to a successful policy in a short time (e.g. one hour).

A second application mode should allow unlimited execution of the learned policy. The agent may not use a simulation created by someone else for the learning process; simulation is of course allowed for evaluating and testing the application. The second difficulty is the low number of interactions with the real scenario, so that measures to increase the efficiency of classic RL approaches have to be used. A suitable visualization of the learning process or the policy or transparent value functions would be helpful in order to clarify the process for visitors and students and to support the program development.

Colloquium: 07.03.2018

Supervisor: Dipl.-Inform. Ingo Boersch, Prof. Dr. Jochen Heinsohn

Download: A1 poster, colloquium lecture, master's thesis

Lecture at the World Usability Day

Thursday November 09, 2017

Colin Christ has been studying computer science at the TH Brandenburg since 2012. He completed his bachelor's degree with a focus on “Intelligent Systems”. In his master’s project he deepened the topic of "Reinforcement Learning (RL)" from the lecture "Artificial Intelligence". Reinforcement learning is a learning paradigm that is increasingly used in robotics. Here, an agent learns an action strategy that is precisely tailored to his situation, for example his body and sensors, through trial and error. New learning algorithms reduce the number of necessary interactions by transferring lived experiences to similar situations and thus lead to a shortened learning process. It now seems possible to use RL in industrial environments.

In his master's thesis, Colin Christ deals with a scenario in which a humanoid robot is supposed to achieve human-defined goals in this way - without explicitly programming how the task can be solved: "Make a wish" programming.

He presented the interim results in an entertaining and stimulating way for discussion at the World Usability Day on this year's topic "Artificial Intelligence" on November 9th, 2017 in the info panel "UX and Design Innovations from Brandenburg" in Berlin.

 

 

Bachelor thesis by Darya Martyniuk

Thursday September 21, 2017

Detection and repair of inconsistencies in a medical ontology

The aim of the bachelor thesis is the investigation of a medical ontology (WNC ontology *) with regard to inconsistencies and the development of an algorithm for the repair of inconsistencies. For this purpose, various inconsistency dimensions in the present ontology were analyzed and a group of inconsistencies were defined, for which a repair is possible with a semi-automatic method. For this purpose, logically contradicting components of the ontology are disambiguated in a dialogue with the domain expert and then correctly modeled.

* The WNC ontology is developed by the Berlin company ID GmbH & Co. KGaA. It depicts the medical terms represented in the Wingert terminology as concepts and the interrelationships between these terms as relations and emerged as the result of the migration of knowledge representation from a semantic network into a modern ontology based on description logic.

Colloquium: 09/21/2017

Supervisor: Prof. Dr.-Ing. Jochen Heinsohn, Dipl.-Inform. Ingo Boersch

Download: A1 poster

Bachelor thesis by Sebastian Fabig

Thursday September 21, 2017

Prediction of dynamic engine processes with long short-term memory neural networks

The aim of the work is to investigate whether and how well LSTM networks are suitable for predicting engine signals.

LSTM networks are recurrent artificial neural networks with a special neuron model. These networks are suitable for forecasting (time series, language processing) or as a generator, especially when the relevant information is in the distant past.

The thesis examines the behavior of the networks on artificial time series and on real signals in the holdout process. The deep learning framework Keras is used on TensorFlow, but without GPU support.

It could be shown that LSTM are able to learn dynamic engine processes and carry out successful prognoses. In the investigated cases in this first pilot project, the prognosis with LSTM delivered similarly good results without any optimization as the models already established at IAV, but it is significantly more complex in terms of training.

Colloquium: 09/21/2017

Supervisor: Dipl.-Inform. Ingo Boersch, Prof. Dr.-Ing. Jochen Heinsohn, Dipl.-Ing. Frank Beyer (IAV GmbH)

Download: A1 poster

Bachelor thesis by Holger Menz

Tuesday, August 08, 2017

Prototypical implementation and test of a restoration algorithm for digital images with motion blur interference

The aim of the work is a software filter to remove the motion blur in photos, which was created by moving the camera during the recording. This blurring disturbance can be modeled in a simplified way as a convolution of the image with a motion kernel and additive noise. If the correct kernel is found, the undisturbed image can be restored in an optimization task.

For this purpose, based on a method published in [Xu10], we developed our own simplified algorithm, which iteratively approaches a suitable kernel by working out the original edges and optimizing. The filter is implemented as a software prototype in Java and inserted into the Fiji image processing software as a plug-in. The plugin is to be used for the restoration of gray value images. The quality of the restoration results and the running time of the resulting filter are analyzed in three test scenarios (two with known kernels, one real image). The procedure was successfully evaluated.

[Xu10] Li Xu, Jiaya Jia: Two-Phase Kernel Estimation for Robust Motion Deblurring. ECCV (1) 2010: 157-170

Colloquium: 08.08.2017

Supervisor: Prof. Dr. sc. techn. Harald Loose, Dipl.-Inform. Ingo Boersch

Download: A1 poster

THB students at the Data Mining Cup 2017

Dynamic Pricing: THB students at the Data Mining Cup 2017

Also this year two teams from the Technical University of Brandenburg took part in the Data Mining Cup of prudsys AG, a student competition for intelligent data analysis. With 202 universities from all over the world, the number of participants reached a new record. The task from the "Dynamic Pricing" field consisted of forecasting a customer transaction for an online pharmacy. For this purpose, real data was available over a period of 3 months, while the following month was to be predicted.

The task is similar to the tasks of previous years, in which it became clear that less the optimization of the learning process than the generation of features in particular contributes to good results. However, this can hardly be automated and is labor-intensive, so that there was a lot to do for the two master's students Mario Kaulmann and Herval Bernice Nganya Nana. Presumably, many teams felt the same, because of the 202 registrations only 66 (!) Teams managed to submit a solution, of which 11 were still invalid due to unusual format requirements. The THB team came 42nd among the 55 valid solutions, suitable for IT specialists. The second submission was unfortunately one minute too late and would have reached 38th place.

Free Will Prediction - Feasibility and Initial Results

Do you act randomly? I know what you will do

Master's project at the NWK18: A person's free will is an urban hypothesis and the content of lively research. The focus is on the question of whether there is free will or whether the human being is controlled by the subconscious. As part of this work, an experiment from the field of human-robot interaction is designed and prepared, which should clarify whether the human attempting to act consciously randomly, but unconsciously falls into a pattern. The preliminary investigation clarifies critical problems and justifies the confidence in the determination based on the prognosis of a time series of human actions.

The presentation for the publication will be given by Vanessa Vogel on May 31, 2017 at the Mittweida University of Applied Sciences.

Vogel, Vanessa; Boersch, Ingo: Free Will Prediction - Feasibility and Initial Results. In: 18th Conference of Young Scientists (NWK), Hochschule Mittweida, 2017 (Scientific Reports No. 1), pp. 341-345. ISSN 1437-7624

Master thesis by Sebastian Busse

Conception and prototypical implementation of a software tool for the dynamic creation of diagnostic reports with the help of ontology-based methods

This work describes an approach according to which pathological reports can be structured and completely created.

The implemented software uses templates from the ICCR (International Collaboration on Cancer Reporting) to create a formal model of the three report types for the creation of endometrial, skin and prostate cancer reports. The document representations achieved are knowledge bases which are formulated in the Web Ontology Language (OWL) and are therefore not only machine-readable, but also machine-understandable. Due to the formally specified semantics of the corresponding format, the reports can be checked for completeness using the HermiT Reasoner. Furthermore, the linking of the modeled report components to external medical knowledge bases such as SNOMED CT, NCIT and PathLex is considered.

The description of the ontology-based procedure and the prototypical implementation of the software tool are intended to show a possible form of representation, according to which diagnostic reports in the field of anatomical pathology can be created and processed precisely and completely digitally, dynamically and using specified structural elements.

Colloquium: 04/10/2017

Supervisor: Prof. Dr.-Ing. Jochen Heinsohn, Dipl.-Inform. Ingo Boersch

Download: A1 poster, master's thesis

Master thesis by Franziska Krebs

Development of a prototypical web application for optimized menu planning using terminological knowledge

The aim of the work is an application for menu planning using ontological knowledge from various sources.

One focus is the selection and integrative networking of suitable sources of knowledge in the form of terminology to describe the requirements for a desired menu plan. This includes, for example, recipes, nutritional information and dietary restrictions. The CTS2 terminology server of Fraunhofer FOKUS should be used.

A second focus is the formalization of the planning problem as well as the selection and executable implementation of a suitable optimization process for multi-target optimization. State-of-the-art technology must be included. The quality of the plans created is evaluated.

The analysis, conception and implementation enable real planning problems in the discourse area to be solved based on the work. The web application runs in an up-to-date browser and allows the terminology to be displayed in a prototypical manner, the input of the planning problem, the parameterization of the planning and the visualization of the results. The difficulty of the work lies in the complexity and in the scope of the partial aspects that have to be solved.

Colloquium: March 24, 2017

Supervisor: Dipl.-Inform. Ingo Boersch, Prof. Dr. rer. nat. Rolf Socher, in cooperation with Fraunhofer FOKUS

Download: A1 poster, master's thesis

Master thesis by Christoph Gresch

Use of data mining to identify and estimate the number of banana plants in an aerial photograph

In this work, banana plants are detected in an orthorectified aerial photo taken with drones and then counted. The task falls within the field of Pattern recognition and should be solved by approaches of machine learning in a data mining process. The particular difficulty lies in the strong overlap of the banana plants in the image, which makes segmentation difficult or even impossible. Likewise, a grid-like arrangement of the plants can no longer be assumed in a developed plantation.

Starting from a number of manually annotated banana templates, features of image points are developed, which should make it possible to separate the centers of the plants from other image points with the help of supervised learning. Various characteristics (color, texture, gray-level co-occurrence matrix) and learning algorithms are systematically examined.

Colloquium: February 27, 2017

Supervisor: Prof. Dr.-Ing. Sven Buchholz, Dr. Frederik Jung-Rothenhäusler (ORCA Geo Services), Dipl.-Inform. Ingo Boersch

Download: A1 poster

Data mining in resistance spot welding. In: Int J Adv Manuf Technol

Wednesday December 21, 2016

International Journal of Advanced Manufacturing Technology

With resistance spot welding, thin sheets are welded to one another at points using high currents in the kiloampere range. The main area of ​​application of the joining process is the construction of motor vehicles and rail vehicles, where these connections must meet safety-critical requirements. Despite intensive research in the field, no sufficiently reliable method is known about the quality of the welded joint non-destructive to determine.

Since the quality of the connection is largely determined by the wear and tear of the electrode caps, the electrodes are milled or replaced with a large safety buffer. We introduce an approach as an essential quality feature - the Point diameter - Estimated sufficiently reliably with forecasting methods on the basis of process variables and thus the electrode life can be increased.

The publication is available in full text for libraries and institutions with a SpringerLink license:

Boersch, Ingo; Füssel, Uwe; Gresch, Christoph; Großmann, Christoph; Hoffmann, Benjamin:
Data mining in resistance spot welding: A non-destructive method to predict the welding spot diameter by monitoring process parameters.

Bachelor thesis by Tobias Meyer

Friday October 14, 2016

Hyperparameter Selection for Anomaly Detection with Stacked Autoencoders - a Deep Learning Application

The aim of the work is to investigate numerical and strategic parameters in the use of autoencoders to detect anomalies in images. The influences of various attitudes must be systematically evaluated and assessed. As a result, a recommendation is to be made to discontinue the procedure for the detection of malaria-infected blood cells. The particular difficulty of the work is the implementation of a systematic search process in an extensive parameter space, the work with real data and the extensive training.

Colloquium: October 14, 2016

Supervisor: Dipl.-Inform. Ingo Boersch, Prof. Dr.-Ing. Sven Buchholz, Dr.-Ing. Christian Wojek (Carl Zeiss AG)

Download: A1 poster

Master thesis by Patrick Rutter

Tuesday, August 02, 2016

Human Robot Interaction using the example of a NAO robot playing tic-tac-toe

The aim of the work is to develop an application that enables a NAO robot to play tic-tac-toe autonomously against a human player. The focus here is on a natural and motivating gaming experience. To this end, it is necessary to develop robust solutions for sub-problems of interaction such as reception and actuation that take this objective into account. A powerful game strategy must be implemented in such a way that both strong and weaker players enjoy the interaction. The application should run autonomously on the robot and be used in the future for other games as well as the NAO versus NAO game. The particular difficulty of the work lies in the design of the interaction and the solving of robotics problems in a real, stochastic world.

The work was divided into the problem areas of game logic, strategy, actuators, image processing and interaction. Game logic and strategy deal with the implementation of the basic gameplay. The actuators are primarily used to implement drawing on the field. In the image processing, the playing field is recorded and evaluated with the help of the robot cameras. In the interaction, a language-based interface with the human opponent as well as an adaptive skill level is implemented.

Colloquium: 08/02/2016

Supervisor: Prof. Dr.-Ing. Jochen Heinsohn, Dipl.-Inform. Ingo Boersch

Download: A1 poster

Master thesis by Benjamin Hoffmann

Modeling of patient-oriented outcomes using data mining methods from data from the treatment process for breast cancer

The aim of the work is to create valid, transparent, predictive models for predicting patient-oriented target variables (poZg), such as the survival of breast cancer patients, from the data of the tumor center. The analysis serves in particular to reveal previously unknown relationships, influencing variables and patterns that can serve to improve the treatment process and can be discussed with doctors. The quality of the results must be suitably assessed on the basis of the available data and evaluated by technical experts (doctors, TZBB).

Associated tasks include: definition of patient-oriented target variables, descriptive and exploratory analysis, determination of relevant features, feature definition, modeling and evaluation.

A second focus is the appropriate patient-oriented visualization of relationships that can be helpful in patient decisions. Particular difficulties in the work are the implementation of the data mining process with real, incomplete, error-prone data and the use of transparent modeling and visualization to gain knowledge for experts and to support decision-making for patients. All software modules should be designed for reusability, also by users at TZBB, Python should preferably be used.

Colloquium: 06/13/2016

Supervisor: Dipl.-Inform. Ingo Boersch, Prof. Dr.-Ing. Sven Buchholz

Download: A1 poster

THB project conference: Robot David plays with adaptive skill levels

Master's student Patrick Rutter demonstrated on June 1st at the project conference of the THB, the capabilities of the NAO robot programmed by him at the TicTacToe game against visitors. With the help of an extended finger, the NAO places its field on a touchscreen and recognizes the movements of the person. During the game he tries to win, but not often in a demotivating way. It adapts to the skill level of the human player and creates an entertaining gaming experience.

The presentation shows an interim status of the master's thesis.


Photo: Patrick Rutter

Franziska Krebs presents her project at the NKIF 2016 at the HAW Hamburg

From May 19 to 21, professors, employees and students from North German universities of applied sciences met for an informal exchange about research and teaching at the 21st North German Colloquium for Computer Science at Universities of Applied Sciences (NKIF 2016) at the HAW Hamburg. For the TH Brandenburg, Franziska Krebs (master's student in computer science) presented her interim results on the SmartMenu project, in which menu planning is carried out based on ontologies and semantic technologies.

Presentation slides: Development of a knowledge network for a clinical meal recommendation system - an example from the THB research / project study

Contacts were made and refreshed, talked shop and discussed.

Photos from the colloquium

Master thesis by Maik-Peter Jacob

Thursday April 07, 2016

Reflection and analysis of the therapy decision in the real treatment process for breast cancer

The aim of the work is to reflect on and analyze the therapy decision in the real treatment process for breast cancer. For this purpose, a normative model, which is obtained from the S3 guideline (LL), is to be compared with the actual conditions, given by epidemiological data from the Tumorzentrum Land Brandenburg e.V.

Related questions are: to what extent has the guidelines been complied with, where are there deviations, how large are the deviations. Then different models and visualizations with data mining methods are to be created on the basis of data. These should reflect the real therapy decision. Important questions for the subsequent reflection and analysis are: where are there overlaps and differences to the LL model, what are the differences, are there other influencing factors than those listed in the guideline.

The particular difficulty of the work lies in the data quality and the complicated application domain.

Colloquium: 04/07/2016

Supervisor: Dipl.-Inform. Ingo Boersch, Prof. Dr. med. Eberhard Beck

Download: A1 poster

The first international workshop experience - the Department of Computer Science and Media makes it possible

Tuesday November 03, 2015

Sebastian Busse presents his project work in Lisbon

Sebastian Busse presents his project work as part of a poster presentation. (Photo: Dennis Wagner)

In the master's project "Artificial Intelligence", Sebastian Busse is working on a system for checking pathological reports for completeness with the help of terminological knowledge. He submitted the published publication [1] to the 2015 LOUHI * workshop in Lisbon. The workshop is part of the EMNLP ** conference.

"Congratulations, your submission has been accepted to appear at the conference."

The Department of Computer Science and Media (FBI) supports students with conference contributions with travel expenses and conference fees in order to promote entry into the conference industry. Sebastian Busse was able to successfully present his work to the scientific discourse in Portugal.

* Sixth Workshop on Health Text Mining and Information Analysis

** Conference on Empirical Methods in Natural Language Processing /

[1] Busse, S. Checking a structured pathology report for completeness of content using terminological knowledge. Proceedings of the Sixth International Workshop on Health Text Mining and Information Analysis. 2015 Sep; 103-108

Bachelor thesis by Jan Dikow

Tuesday September 15, 2015

Dimension reduction of categorical data for the generation of thematic maps

The mapegy company generates various visualizations based on several data sources such as patent data and scientific publications for their web-based analysis and visualization software mapegy.scout. One of the visualizations is a Patent mapwhich grouping the patents on the basis of the user-dependent input (cluster analysis) and shows these groups on a map (dimension reduction) so that similar patents are close together and different ones are further apart. This process should be fundamentally revised so that

  1. Different types of documents (also e.g. news and scientific publications) can be processed based on their assignment to certain categories,
  2. the process becomes more scalable and faster overall,
  3. first results are provided quickly (e.g. through a preview, pre-processes or sampling),
  4. an output data model is created that enables various representations in the front end.

A GHSOM (Growing Hierarchical Self-Organizing Map) selected whose individual sub-maps consist of a number of neuron models that adapt to the training data and thus realize clustering and dimension reduction at the same time.

Colloquium: 15.09.2015

Supervisor: Dipl.-Inform. Ingo Boersch, Uwe Kuehn, M.Sc. (mapegy GmbH, Berlin)

Download: A1 poster

Master thesis by Andy Klay

Monday September 14, 2015

Realization of a tic-tac-toe-playing NAO robot by means of automatic learning of the game strategy

The aim of the work is to develop an application that lets a NAO robot play tic-tac-toe against a human. In the future, a game between NAO robots should also be possible. One focus of the work is the suitable implementation of a learning process with which the application learns a game strategy, e.g. with reinforcement learning. The learning process should be evaluated by measuring the skill level. A display of the current skill level is desirable.

The application should be designed in a modular manner so that components can be exchanged or expanded easily.It should be possible to implement a related game, such as 4x4 Tic-Tac-Toe, in which essentially only the game-dependent components (such as game rules, situation recognition, move execution and test opponents) are modified. The game control and learning module components should be as independent as possible from the specific game.

A sub-task consists in recognizing the game situation with the help of image processing. The position of the playing field relative to the robot can be assumed to be static and known. It is to be suitably defined in the context of the work. A sensible, simple interface must be implemented to control the actuators, taking the NAO platform into account. The application and results are to be evaluated in a suitable manner.