Constraint-based Models and Algorithms for Self-Diagnosis and -Planning
|
Our group develops constraint-based modeling techniques and algorithmic methods at the intersection of artificial intelligence, combinatorial optimization and embedded system engineering, in order to support and automate the programming of "self-aware" systems that can autonomously diagnose, test, repair, and optimize themselves by reasoning from models of their normal and faulty behavior.
|
|
Research Topics
- Constraint-Based Diagnosis and Planning
- Constraint Optimization Algorithms
- Model-Based Programming
- Model-Based Reasoning
- Model-Based Systems
- Probabilistic Models
- Logical Models
- Automatic Planning
- Test Generation
- Failure Diagnosis
Application Domain
- Cognitive Factory
- Intelligent Kitchen
Team
Group Details
How to represent the behavior of complex technical (or natural) systems in a computer model? How to get from quantitative data (e.g. obtained from sensor measurements) to a high-level, qualitative understanding of the system (e.g. its relevant states and possible transitions)? How to exploit such models to track the state of the system and drive it towards desired states? How to make the models concise enough and the algorithms that operate on them fast enough to equip everyday devices (e.g. cars) with model-based self-diagnosis, self-repair, and self-optimization capabilities?
We develop algorithmic and model-theoretic foundations for "self-aware" systems that can autonomously diagnose, test, repair, and optimize themselves by reasoning from models of their normal and faulty behavior. The models are expressed as constraints that describe feasible states of system components, and (probabilistic or cost-based) transitions between these states. The main challenge is to develop algorithms and representations that allow for efficiently reasoning through a large space of component interactions to quickly identify most likely current states (monitoring and diagnosis) and generate least-cost actions to drive a system towards desired states (planning and reconfiguration). While this is very hard (NP-complete) in general, real-world instances are likely to have structural features in their underlying constraint model that make these tasks tractable and even feasible for on-line, embedded applications. Our approach is therefore to adapt and combine techniques from constraint programming, combinatorial optimization, graph theory, and probabilistic reasoning to exploit model structure and obtain practically effective and scalable reasoning strategies. We seek to apply our methods in the context of programming intelligent embedded systems (such as fault-robust automotive subsystems and flexible manufactirung systems) and software systems (such as self-configuring web services).
