Foundations of Artificial Intelligence (FAI) Group
divider line

   Home    People    Research    Projects    Software    Teaching    Past Student Projects    Publications

Explaining the Space of Plans (AFOSR, 07/18 -- 06/23)

This project is conducted jointly with Daniele Magazzeni at King's College London. It is funded by AFOSR (U.S. Air Force Office of Scientific Research).

Explainability of AI methods is one of the grand challenges at this time. Model-based AI lends itself naturally to this purpose, as it takes decisions based on explicit reasoning about world behaviour as captured in the model. The difficulty then lies in actually making such reasoning -- enumerating vast spaces of alternate possibilities -- amenable to human users. Our central thesis in this project is that this can be naturally done in terms of explaining the space of plans, pointing out the most relevant plan properties and their dependencies.

In the project, we capture such dependencies in the form of plan-property dependency networks (PDN). We develop suitable concepts, and technologies for inferring PDNs, for a range of planning frameworks including classical planning, temporal planning, and oversubscription planning.

Controlling Information Density in Discourse Generation (DFG, 07/18 - 06/22)

This project is part of the DFG collaborative research center (Sonderforschungsbereich) Information Density and Linguistic Encoding. The project is conducted jointly with Alexander Koller. It is funded by DFG (German Research Foundation/Deutsche Forschungsgemeinschaft).

The FAI part of this project investigates planning methods for natural language generation, specifically for the automtic generation of instructions, using Minecraft as a benchmarking environment. Some background information on FAI's research in language generation is given on our Research page. As a whole, the project investigates ways of modeling the interaction between instruction giver (IG) and instruction follower (IF) as a rational speech act, where the IG attempts to obtain compact instructions while limiting the risk that the IF misunderstands these. These models will be operationalized through text-planning methods planning the overall instructions, and surface realization methods planning individual sentences.

Optimizing Planning Domains (DAAD, 01/17 -- 12/18)

This project is conducted jointly with Carlos Areces at Universidad Nacional de Cordoba, Argentina. It is funded by DAAD (German Academic Exchange Service/Deutscher Akademischer Austauschdienst), and supports travel for research visits between the two locations.

The project investigates methods for optimizing the input to planning systems, i.e., the models (the planning domains) to be solved. There is usually a multitude of ways in which the same problem can be modeled, and these differences have a huge impact on performance. Thus domain optimization is important. Yet that topic has traditionally been neglected in the literature. In this project, we start from our work at ICAPS'14 on action schema splitting to reduce pre-processing effort, and take this direction further, to other domain transformations, and to the question when two distinct domain models can be considered to "model the same problem".

Simulated Network-Penetration Testing (BMBF, 08/16 -- 07/20)

This project is conducted jointly with Michael Backes at CISPA. It is funded by BMBF (Federal Ministry of Education and Research/Bundesministerium fuer Bildung und Forschung).

The project investigates models and methods for simulated penetration testing (see our Research page). It also considers advanced methods like what-if analyses, and applications to security analysis beyond company networks.

Star-Topology Decoupled State Space Search (DFG, 05/16 -- 04/19)

This project is funded by DFG (German Research Foundation/Deutsche Forschungsgemeinschaft).

The project investigates star-topology decoupling (see our Research page) and its combination with other search methods. It initially explores the application to model checking.

Critically Constrained Planning via Partial Delete Relaxation (DFG, 01/15 -- 12/16 and 01/17 -- 06/19)

This project is funded by DFG (German Research Foundation/Deutsche Forschungsgemeinschaft).

The project investigates partial-delete relaxation methods and their properties (see our Research page), with a special focus on resource-constrained planning where the goal needs to be attained subject to fixed resource budgets.

Data protection