Model-based problem solving

1. Knowledge compilation for diagnostic tasks: KCII-DST

Deep knowledge and knowledge compilation is one of the advanced research topics aiming at overcoming the brittleness of the current expert systems. Deep knowledge is fundamental knowledge of domains which could give the expert systems high flexibility and capability of problem solving. A knowledge compiler has a powerful reasoning mechanism using deep knowledge and produces shallow knowledge by chunking the inference process which will be easily used in similar situations in future problem solving.

The results of knowledge compilation research includes a domain-specific tool named KCII-DST which helps the user build a diagnostic system of mechanical artifacts. The only thing the user has to do is to give information about the structure and function of his/her device through a graphical interface. The diagnostic system first tries to diagnose the device by using the shallow knowledge. When it fails, it invokes the deep engine and diagnosis is done using the deep knowledge. When the deep diagnosis is successful, the reasoning process is compiled to obtain shallow knowledge.

2. Qualitative reasoning and knowledge compilation: KCIII

Although KCII-DST generates diagnostic rules for several types of mechanical systems, an application of KCII-DST to a nuclear plant showed that it sometimes generates ambiguous results due to its complex topologies, which motivates us to develop KCIII. In the research on qualitative reasoning, little is known about how the methodology can be applied to more complex systems. Many applications of qualitative reasoning to complex systems suffer from ambiguous results.

In this research project, we concentrate on the following three issues concerning diagnostic ES's;

The new features of KCIII realized for satisfying these requirements are causal specifications, utilization/compilation of global knowledge and two time counters. The causal specifications represent a component's local causal properties satisfying the principles for reusability and composability. A model of a system is represented by combining a set of local component models and global knowledge derived from general properties of the physical entity. This allows for reusable knowledge which is easy to describe. The reasoning engine has two time counters for polynomial equations and uses heuristics for negative feedback. They contribute to generating intuitive causal ordering of complex behavior including inter-components' negative feedback and unambiguous reasoning results.

KCIII has been successfully applied to a heat transportation system of a nuclear power plant whose model has 47 parameters and 27 equations. Reasoning results were unambiguous and matched those obtained by domain experts.

Reports:

1995

MizLab Homepage