Maier, Paul
Researcher
PhD student
Informatics
|
Table of ContentsResearch TopicsPublications |
Students: check out open projects (SEP/BA/MA/DA) and open Hiwi positions in our group here.
I'm interested in model-based reasoning in embedded systems, specifically automatic constraint-based diagnosis, planning and configuration. Model-based diagnosis and planning provide means to handle errors and unanticipated situations in today's increasingly complex embedded systems, model-based configuration makes it easier to reconfigure or restructure components of such systems.
In general, I'm dealing with the problem of belief state update in embedded systems. Computing the exact belief state is infeasable due to its exponential complexity, so the aim is to find a good approximation to the belief state. Additionally, since ressources in embedded systems tend to be extra sparse it is critical that they are optimaly utilized.
In my thesis I aim to address two problems that arise in this context: 1) a tighter integration of diagnosis and planning, which can be achieved using soft constraints as common representation for both of them, framing them as constraint optimization problems (COP), and 2) a more flexible belief state approximation which can be developed based on this representation using adaptive abstraction.
In an embedded system context, sets of feasible system states as well as state transitions can be specified by constraints. Soft constraints are an extension to constraints which allow to differentiate feasible states according to some kind of preference, such as transition probabilities or costs. Soft constraints therefore allow for very flexible modeling and easy extension (by simply adding constraints). Within this context they are used to represent probability distributions over sets of system states and sets of state transitions. A complete system model is then represented as constraint-based HMM (Hidden Markov Model).
Solving COPs is NP-hard, but it can be made feasible through approximations. One possibility to approximately solve a COP is to employ domain abstraction: variable values are aggregated into abstract values. This means that less variable distinctions have to be considered, which greatly reduces the problems size. The number of values which get abstracted can be varied,e.g., adapted to a size bound (the more values are abstracted, the less space is needed for storing the abstract domain). I aim to employ these kinds of adaptive abstraction methods to develop a more flexible belief state approximation.
Interesting application domains are factory plants such as the cognitive factory or the demo plant at ITM (used in the practical "Praktikum Simulationstechnik"), and the intelligent kitchen. Currently, I focus on the ITM demo plant as an example scenario. Below is a photograph of the demo plant.
Research Topics
- Constraint-Based Diagnosis and Planning
- Model-Based Reasoning
- Constraint Optimization Algorithms
- Model-Based Programming
- Probabilistic Reasoning
Education
- 2001-2007 Computer Science (Diploma) at the University of Passau
- Diploma thesis: "Content-Based Image Retrieval: Review and Benchmarking"
Students
Conferences:
- CPAIOR'08
- AAAI'08
- KI'09
- ETFA'09
- PHM'09
Misc:
- Toolbar is a open source suite of constraint optimisation algorithms (based on weighted CSP) developed at INRIA. I integerated my algorithm BEDA into this suite. It can be downloaded at:http://mulcyber.toulouse.inra.fr/gf/project/toolbar/
- Livingstone 2 is an open source diagnosis engine developed at the model-based diagnosis and recovery group at NASA. Download it at http://opensource.arc.nasa.gov/project/livingstone2/
