One of the main problems in the design of intelligent systems is the so-called "knowledge engineering bottleneck": human experts master their respective tasks, but are nevertheless unable to communicate their problem solving techniques in a way that would allow to directly transfer their expert knowledge into a computer program. The knowledge engineer is the mediator between man and machine in this process.

The "Knowledge Engineering" group concentrates on the acquisition of explicit, formalizable knowledge from sources that contain relevant information in implicit or not directly accessible form. Its methodological focus is on the use of techniques from machine learning and data mining for knowledge acquisition by analysis of existing data or text collections, by interaction with human experts, or by experimentation and simulation. The applicability of the employed methodologies ranges from cognitive science problems to industrial applications in the areas of data or web mining.

Our main research goals are summarized below, a list of current projects and our publications are available on separate pages.

Machine Learning and Data Mining

Machine learning is concerned with techniques and methodologies for enabling programs to autonomously perform changes in their behavior based on experience. While the early days of the field focused on cognitive models of learning, it subsequently developed more and more towards the goal of finding patterns and regularities in existing data collections. Typical machine learning algorithms are classification, regression and clustering algorithms. As such, it developed into one of the core research areas in data mining. Just like the gold miner, who mines a mountain for nuggets of gold, the data miner searches large databases for hidden nuggets of knowledge.

Inductive Rule Learning

Our research is not restricted to a particular learning scenario. Instead, we focus on the learning of simple and understandable forms of knowledge, most notably IF-THEN rules and rule sets, in a wide variety of basic research and application areas. On the one hand, our basic research has contributed to a better understanding of the fundamentals of inductive rule learning, but it has also equipped us with a set of tools that we can put to practice in applied research projects. Our goal is to become a recognized competence center for rule learning.

Learning from Preferences

In many applications (e.g., recommender systems), the training information is stated in the form of preferences between a set of object or labels. In addition, many conventional learning problems, such as classification, multilabel classification, and ranking can be re-formulated as preference learning problems. In our work, we address this problem with pairwise classification, a technique that learns a separate theory for each pair of labels. Despite its seeming complexity, we showed in several scenarios the technique is not more expensive than alternative proposals. In this area we have a strong collaboration with the Intelligent Systems Group of the University of Paderborn.

Multilabel Classification

Conventional multiclass classification only allows a mapping of a training example to exactly one class.  In Multilabel Classification problems, each object or document can be mapped to an arbitrary number of labels. A typical example of multilabel classification is the genre categorization of stories. Our work is mainly based on approaches that tranforms the problem into a preference learning problem. In this area, we have an on-going co-operation with the Machine Learning and Knowledge Discovery Group of the Aristotle University of Thessaloniki.

Information Extraction

Information Extraction is the task of identifying certain types of information in documents, usually text documents. A common task is, e.g., Named Entity Recognition which aims at identifying names of people, organizations, places, temporal expressions etc. in texts. We focus on the use of machine learning techniques in order to learn useful extraction patterns.

Web Mining

Web Mining is a young discipline which tries to employ machine learning and data mining techniques for facilitating a semantic processing of the abundance of documents on the Web. The key difference to data mining lies in the fact that the data material is not in the form of structured relational or multi-relational databases, but consists of unstructured text or semi-structured documents. In particular in the latter case, the structure of HTML-documents on the hand and their topology on the other hand may provide important pieces of information that we want to utilize with suitable learning architectures.

Semantic Web

The Semantic Web is a vision coined by Tim Berners-Lee, referring to a web of information which can be processed by both men and machines (unlike the "traditional" World Wide Web, in which machines can hardly access the information contained therein), currently manifesting as Linked Open Data. We are particularly interested in applying machine learning methods to the Semantic Web. Example applications are: learning ontologies and rules on Linked Open Data, using Web Mining for automatically creating Semantic Web documents from traditional HTML, and using information from the Semantic Web for improving machine learning algorithms.

Game Playing

Since the early days of cognitive science and artificial intelligence, game playing has been a popular study domain. For example, Nobel laureate and AI pioneer Herbert Simon has shown in the domain of chess that expert players differ from beginners not so much in their ability to look ahead, but much more in the way in which they are able to re-formulate a problem in terms of known entities (so-called chunks), which allows them to structure and simplify the problem. One goal of our group is to use modern machine learning and data mining techniques for bridging the gap between knowledge-based and cognitive approaches and the data- and computation-intensive techniques that currently predominate in game playing and many other areas of AI. Our research is wide-spread and deals with classic games such as chess and go as well as with AI from commercial games. An interesting game from the point of view of AI which we have been investigating in recent times is poker. Its peculiarities in game playing are the multiplayer setting, the incomplete information and randomness. We have successfully participated in the 2008 (2nd place) and 2009 (3rd place) Computer Poker competitions in the multi-player leagues.

Predictive and condition based Maintenance

The domain of Predictive Maintenance is about methods and technologies helping to estimate the future condition of a system that is subject to degradation caused by normal operation. The goal of those techniques is to minimize the long-term cost of maintenance activities while guaranteeing certain quality constraints. To achieve this goal, condition monitoring is used to provide continuous information on the status of the system and its components. By incorporating the status information into a model, degradation processes can be anticipated. In contrast to preventive or corrective maintenance, a predictive method can reduce the cost of maintenance by avoiding severe failures. Advanced business scheduling processes for maintenance activities can be implemented to further reduce cost and save time by improving the handling of necessary repairs.

small ke-icon

Knowledge Engineering Group

Fachbereich Informatik
TU Darmstadt

S2|02 D203
Hochschulstrasse 10

D-64289 Darmstadt

Telefon-Symbol+49 6151 16-21811
Fax-Symbol +49 6151 16-21812

A A A | Drucken | Impressum | Sitemap | Suche | Mobile Version
zum Seitenanfangzum Seitenanfang