Thomas A. Runkler’s research while affiliated with Technical University of Munich and other places

What is this page?


This page lists works of an author who doesn't have a ResearchGate profile or hasn't added the works to their profile yet. It is automatically generated from public (personal) data to further our legitimate goal of comprehensive and accurate scientific recordkeeping. If you are this author and want this page removed, please let us know.

Publications (280)


Visualization of Preference Matrices for Labeled Objects
  • Chapter

January 2025

Thomas A. Runkler


TEA: Trajectory Encoding Augmentation for Robust and Transferable Policies in Offline Reinforcement Learning

November 2024

·

5 Reads

In this paper, we investigate offline reinforcement learning (RL) with the goal of training a single robust policy that generalizes effectively across environments with unseen dynamics. We propose a novel approach, Trajectory Encoding Augmentation (TEA), which extends the state space by integrating latent representations of environmental dynamics obtained from sequence encoders, such as AutoEncoders. Our findings show that incorporating these encodings with TEA improves the transferability of a single policy to novel environments with new dynamics, surpassing methods that rely solely on unmodified states. These results indicate that TEA captures critical, environment-specific characteristics, enabling RL agents to generalize effectively across dynamic conditions.


Classification

November 2024

·

3 Reads

Classification is supervised learning that uses labeled data to assign objects to classes. We distinguish false positive and false negative errors and review numerous indicators to quantify classifier performance. Also pairs of indicators are often considered to assess classification performance. We illustrate this with the receiver operating characteristic and the precision recall diagram. Several different classifiers with specific capabilities and limitations are presented in detail: the naive Bayes classifier, linear discriminant analysis, the support vector machine (SVM) using the kernel trick, nearest neighbor classifiers, learning vector quantification, and hierarchical classification using decision trees.


Regression

November 2024

·

19 Reads

Regression estimates functional dependencies between features. Linear regression models can be efficiently computed from covariances but are restricted to linear dependencies. Substitution allows to identify specific types of nonlinear dependencies by linear regression. Robust regression finds models that are robust against outliers. A popular class of nonlinear regression methods are universal approximators. We present two well-known examples for universal approximators from the field of artificial neural networks: the multilayer perceptron and radial basis function networks. Universal approximators can realize arbitrarily small training errors, but cross-validation is required to find models with low validation errors that generalize well on other data sets. Feature selection allows us to include only relevant features in regression models leading to more accurate models.



Correlation

November 2024

·

8 Reads

Correlation quantifies the relation between features. Linear correlation methods are robust and computationally efficient but detect only linear dependencies. Nonlinear correlation methods are able to detect nonlinear dependencies but need to be carefully parametrized. As an example for nonlinear correlation we present the chi-square test for independence that can be applied to continuous features using histogram counts. Nonlinear correlation can also be quantified by the cross-validation error of regression models. Correlation does not imply causality. Spurious correlations may lead to wrong conclusions. If the underlying features are known, then spurious correlations may be compensated by partial correlation methods.


Data Preprocessing

November 2024

In real world applications, data usually contain errors and noise, need to be scaled and transformed, or need to be collected from different and possibly heterogeneous information sources. We distinguish deterministic and stochastic errors. Deterministic errors can sometimes be easily corrected. Inliers and outliers may be identified and removed or corrected. Inliers, outliers, or noise can be reduced by filtering. We distinguish many different filtering methods with different effectiveness and computational complexities: moving statistical measures, discrete linear filters, finite impulse response, infinite impulse response. Data features with different ranges often need to be standardized or transformed.


Data Visualization

November 2024

·

38 Reads

Data can often be very effectively analyzed using visualization techniques. Standard visualization methods for object data are plots and scatter plots. To visualize high-dimensional data, projection methods are necessary. We present linear projection (principal component analysis, Karhunen-Loève transform, singular value decomposition, eigenvector projection, Hotelling transform, proper orthogonal decomposition, multidimensional scaling) and nonlinear projection methods (Sammon mapping, auto-encoder). Data distributions can be estimated and visualized using histogram techniques. Periodic data (such as time series) can be analyzed and visualized using spectral analysis (cosine and sine transforms, amplitude and phase spectra).


Clustering

November 2024

·

19 Reads

Clustering is unsupervised learning that assigns labels to objects in unlabeled data. When clustering is performed on data that possess class labels, the clusters may or may not correspond with these classes. Cluster partitions may be mathematically represented by sets, partition matrices, and/or cluster prototypes. Sequential clustering (single linkage, complete linkage, average linkage, Ward’s method, etc.) yields hierarchical cluster structures but is computationally expensive. Partitional clustering can be based on hard, fuzzy, possibilistic, or noise clustering models. Cluster prototypes can have different shapes such as hyperspheres, ellipsoids, lines, circles, or more complex shapes. Relational clustering finds clusters in relational data, often enhanced by kernelization. Cluster tendency assessment finds out if the data possess a cluster structure at all, and cluster validity measures help identify the number of clusters or other algorithmic parameters. Clustering can also be done by heuristic methods such as self-organizing maps.


Citations (36)


... This approach also reduces the amount of data transmitted over networks, reduces the load on servers, and enables the development of more complex sensor systems in the future. While TinyML offers promising benefits such as reduced latency, improved real-time processing, and enhanced security [14], its widespread adoption is hindered by several limitations. Current approaches require manual tuning and optimization, limiting scalability, and ultra-low power consumption limits model complexity, compromising accuracy and robustness. ...

Reference:

Combinative model compression approach for enhancing 1D CNN efficiency for EIT-based Hand Gesture Recognition on IoT edge devices
On-device Online Learning and Semantic Management of TinyML Systems
  • Citing Article
  • May 2024

ACM Transactions on Embedded Computing Systems

... Meta-learning has proved to accelerate model adaptation to arbitrary labels by allowing fine-tuning over small datasets when faced with previously unseen tasks [7]. The adoption of meta-learning in FL concentrates on what is termed as 'perfect setups', which are challenging to implement in real-world applications. ...

TinyReptile: TinyML with Federated Meta-Learning
  • Citing Conference Paper
  • June 2023

... Anna Himmelhuber suggested that to enhance the usefulness of alerts for analysts, symbolic and subsymbolic methods such as knowledge graphs can be used to improve the explainability and quality in modern industrial systems. By integrating symbolic reasoning with subsymbolic data-driven approaches, the research aims to provide actionable insights and facilitate effective responses to cyber threats [40]. While not directly focused on the combination of symbolic and substantive methods, the framework provided by Masike Malatji dedicates valuable insights into the broader AI-driven cybersecurity landscape. ...

Detection, Explanation and Filtering of Cyber Attacks Combining Symbolic and Sub-Symbolic Methods
  • Citing Conference Paper
  • December 2022

Anna Himmelhuber

·

Dominik Dold

·

Stephan Grimm

·

[...]

·

Thomas Runkler

... For example, RDF enables to represent the information that "There is a CO 2 Sensor located in a Space identified by IFC GUID 1 https://www.buildingsmart.org/standards/bsi-standards/industry-foundation-classes/ 2 https://project-haystack.org/ 3 https://www.w3.org/TR/wot-thing-description11/ 4 https://www.w3.org/2019/wot/td 5 https://www.w3.org/TR/rdf-primer/ "0KLkXPBfvES9D1y7EjijkE" on the 9th Floor of the Atlas Building on TU Eindhoven campus Site" (Chamari et al., 2022). The use of the above-mentioned ontologies has been extensively demonstrated in various applications such as knowledge-based fault detection in heating, ventilation and air-conditioning (HVAC) systems (Delgoshaei et al., 2017), portable data-driven applications (Balaji et al., 2018), Building Information Model (BIM) and sensor data integration (Chamari et al., 2022), management of on-device IoT applications (Ren et al., 2022), etc. Besides devices and systems, recent literature also shows attempts to model software artifacts using ontologies. ...

Towards Semantic Management of On-Device Applications in Industrial IoT
  • Citing Article
  • November 2022

ACM Transactions on Internet Technology

... In the literature, numerous proposals have been introduced to tackle the problem of mining gradual itemsets or its variants from numerical data, e.g. [8][9][10][14][15][16][17][18][19][20][21][22]. One of the key challenges associated with mining gradual patterns approaches is the exponential combination space that needs to be explored, compounded by the need of handling a huge number of patterns, potentially of exponential size. ...

A metaheuristic approach for mining gradual patterns
  • Citing Article
  • Full-text available
  • November 2022

Swarm and Evolutionary Computation

... Ren et al. introduced the Semantic Low-Code Engineering for ML Applications (SeLoC-ML) framework in order to assist non-experts in rapid development of ML applications in the IIoT by utilizing semantic web technologies [127]. SeLoC-ML helps to cope with the compatibility challenges in integrating heterogenious non-standardized models. ...

SeLoC-ML: Semantic Low-Code Engineering for Machine Learning Applications in Industrial IoT
  • Citing Chapter
  • October 2022

Lecture Notes in Computer Science

... Model-free RL continuously adjusts its approach based on real-time feedback. Also, for microcontrollers with less programming memory, this model-free RL algorithm can be used due to low memory requirements as compared to other supervised ML algorithms [35,36]. ...

Comparing Model-free and Model-based Algorithms for Offline Reinforcement Learning

IFAC-PapersOnLine

... Particularly, the authors proposed an RNN-based GAN model that is trained jointly with an encoder-decoder network. Recently, Arnout et al. [20] introduced Classspecific Recurrent GAN (CLaRe-GAN) by conditioning the generator on auxiliary input comprising class-specific and class-independent properties. Specifically, the model consists of two encoders, one for each kind of information (inter-and intra-class characteristics), a shared-latent-space assumption, and a class discriminator that discriminates across latent vectors to extract class-specific features. ...

ClaRe-GAN: Generation of Class-Specific Time Series
  • Citing Conference Paper
  • December 2021