Citation Link: https://nbn-resolving.org/urn:nbn:de:hbz:467-3807
Transformation of graphical models to support knowledge transfer
Source Type
Doctoral Thesis
Author
Subjects
Graphical Models
Decision Networks
Framework
Product Lifecycle Management
Condition Monitoring
DDC
620 Ingenieurwissenschaften und Maschinenbau
GHBS-Clases
Issue Date
2008
Abstract
Human experts are able to flexible adjust their decision behaviour with regard to the respective situation. This capability pays in situations under limited resources like time restrictions. It is particularly advantageous to adapt the underlying knowledge representation and to make use of decision models at different levels of abstraction. Furthermore human experts have the ability to include uncertain information and vague perceptions in decision making.
Classical decision-theoretic models are based directly on the concept of rationality, whereby the decision behaviour prescribed by the principle of maximum expected utility. For each observation some optimal decision function prescribes an action that maximizes expected utility. Modern graph-based methods like Bayesian networks or influence diagrams make use of modelling. One disadvantage of decision-theoretic methods concerns the issue of complexity. Finding an optimal decision might become very expensive. Inference in decision networks is known to be NP-hard. This dissertation aimed at combining the advantages of decision-theoretic models with rule-based systems by transforming a decision-theoretic model into a fuzzy rule-based system. Fuzzy rule bases are an efficient implementation from a computational point of view, they can approximate non-linear functional dependencies and they are also intelligible. There was a need for establishing a new transformation process to generate rule-based representations from decision models, which provide an efficient implementation architecture and represent knowledge in an explicit, intelligible way. At first, an agent can apply the new parameterized structure learning algorithm to identify the structure of the Bayesian network. The use of learning approaches to determine preferences and the specification of probability information subsequently enables to model decision and utility nodes and to generate a consolidated decision-theoretic model. Hence, a transformation process compiled a rule base by measuring the utility loss as approximation measure. The transformation process concept has been successfully applied to the problem of representing condition monitoring results for a rotation spindle.
Classical decision-theoretic models are based directly on the concept of rationality, whereby the decision behaviour prescribed by the principle of maximum expected utility. For each observation some optimal decision function prescribes an action that maximizes expected utility. Modern graph-based methods like Bayesian networks or influence diagrams make use of modelling. One disadvantage of decision-theoretic methods concerns the issue of complexity. Finding an optimal decision might become very expensive. Inference in decision networks is known to be NP-hard. This dissertation aimed at combining the advantages of decision-theoretic models with rule-based systems by transforming a decision-theoretic model into a fuzzy rule-based system. Fuzzy rule bases are an efficient implementation from a computational point of view, they can approximate non-linear functional dependencies and they are also intelligible. There was a need for establishing a new transformation process to generate rule-based representations from decision models, which provide an efficient implementation architecture and represent knowledge in an explicit, intelligible way. At first, an agent can apply the new parameterized structure learning algorithm to identify the structure of the Bayesian network. The use of learning approaches to determine preferences and the specification of probability information subsequently enables to model decision and utility nodes and to generate a consolidated decision-theoretic model. Hence, a transformation process compiled a rule base by measuring the utility loss as approximation measure. The transformation process concept has been successfully applied to the problem of representing condition monitoring results for a rotation spindle.
File(s)![Thumbnail Image]()
Loading...
Name
holland.pdf
Size
5.6 MB
Format
Adobe PDF
Checksum
(MD5):983c5ab5f866b6d12358147089602e6e
Owning collection