The Dependable Intelligent Software Lab (DISL) has opened many directions for research, all within the scope of creating dependable AI systems. Students interested in conducting state-of-the-art research within this field, are cordially invited to write their final thesis (Bachelor/Master) on any of the suggested topics. Also, check out the offerings by other groups participating in DISL. Feel free to also contact us with your own related ideas, or in case you prefer to do a research project instead.

Defining Robustness in HTN Planning

Description: Hierarchical task network (HTN) planning has long been used in various application fields to generate executable action sequences to accomplish some tasks. However, real-world environments are typically dynamic and not all plans are equally likely to succeed in such a setting. We would like to favor plans that are more likely to succeed when exposed to change in the environment, and call this ability the robustness of a plan. But various other definitions of robustness, as well as related concepts, exist. Based on prior work in classical planning, this thesis should discuss possible definitions in order to find a definition of robustness applicable to HTN planning that can be intuitively understood using existing example problems.

Contact:  Tobias Schwartz

Robust HTN Planning via CSPs

Description: Hierarchical task network (HTN) planning problems have been tackled with various approaches, from standard search procedures to complex SAT encodings. The focus of new techniques is usually to improve the current state-of-the-art in terms of faster solving speed or the ability to solve more complex problem instances. In contrast, we are interested to find solutions that are robust when executed in real-world settings. For this purpose, we proposed to encode HTN planning problems as constraint satisfaction problems (CSPs), for which measures of robustness have already been studied. There are at least two potential thesis topics: (1) Theoretically improving the existing encoding such that off-the-shelf solvers can be used to find robust plans; (2) Practically adapting a current CSP solver to handle the encoding that we already have.

Contact:  Tobias Schwartz

Repairing Spatial Knowledge bases in the Web of Data

Description: With the ever-increasing enrichment of the Semantic Web with geospatial data, it is often the case that, for example, the geometries of geographical objects are not captured correctly due to contradictory data of different sources. Thus, we can obtain inconsistent spatial information when extracting topological relations from such geometries (e.g., two overlapping regions may be stated to be identical to a third region, which is impossible as they would also have to be identical to each other if that was the case). In this thesis we will explore ways of repairing such inconsistent knowledge bases in a practical manner, and investigate possible trade-offs between obtaining repairs that are good and obtaining repairs in a manner that is fast.

Contact: Michael Sioutis

A Portfolio Solver for Qualitative Constraint-based Reasoning

Description: The prevalence of multicore architectures has led a substantial amount of research effort to the development of portfolio solvers, i.e., meta-solvers that run n > 1 constituent solvers, which can be instances of a single solver with different settings or several different solvers of the same type on the same problem in parallel; if any of the solvers succeeds in finding a solution all the solvers are terminated. Such solvers have been extensively used in the SAT community to great success, but, to the best of our knowledge, there does not exist any similar approach in the field of qualitative constraint-based reasoning, which in an area of study of infinite-domain constraint satisfaction problems. In this thesis we will explore possible portfolio-based approaches for qualitative constraint-based reasoning, and investigate measures and techniques pertaining to the diversification of a portfolio (each solver should contribute sufficiently to the portfolio performance).

Contact: Michael Sioutis

Guarding ML Classification with Ontologies

Description: In several machine learning (ML) tasks, classifiers have to be learnt that assign a label from a finite set of labels to an entity. Tasks that require multiple labels to be associated to an entity are particularly challenging. Often, some background information about the semantics of labels is available in form of an ontology, for example, in a movie genre classification task we may assume 'horror' and 'family' to be mutually exclusive concepts. Current ML classifiers are unable to exploit rich background information for improving labelling accuracy and preventing wrong and potentially dangerous classifier decisions. We have developed a guarded ML learning technique for multi-label, multi-class, and zero-shot learning, which does not violate any constraints imposed. Our method is based on an adaption of support vector machine (SVM) learning. In the proposed Master thesis, we invite you to explore variations of the outlined method in order to improve accuracy of this safe ML method. 

Contact:  Diedrich Wolter