Knowledge-Based Model Construction (KBMC) has generated a lot of attention
due to its importance as a technique for generating probabilistic or decision-theoretic
models whose range of applicability in AI has been vastly increased. However, no
one has tried to analyze the essential issues in KBMC, to determine if there exists
a general efficient KBMC method for any problem domain, or to y identify the
fruitful future research on KBMC. This research presents a unified framework for
comparative analysis of KBMC systems identifying the essential issues in KBMC,
showing that there is no such general efficient KBMC method, and listing the fruitful
future research on KBMC.
This thesis then presents a new KBMC mechanism for hierarchical diagnosis and
repair. Diagnosis is formulated as a stochastic process and modeled using influence
diagrams. In the best case using an abstraction hierarchy in problem-solving can
yield an exponential speedup in search efficiency. However, this speedup assumes
backtracking never occurs across abstraction levels. When this assumption fails,
search may have to consider different abstract solutions before finding one that can be
refined to a base solution, and, therefore, search efficiency is not necessarily improved.
In this thesis, we present a decision model construction method for hierarchical
diagnosis and repair. We show analytically and experimentally that our method
always yields a significant speedup in search efficiency, and that hierarchies with
smaller branching factors yield more significant efficiency gains.
This thesis employs two causal pathways (functional and bridge fault) of domain
knowledge in device trouble shooting, preventing either whole class of faults we will
never be able to diagnose. Each causal pathway models the knowledge of adjacency
and behavior within the corresponding interaction layer. Careful search of causal
pathways allows us to restrict the search space of fault hypotheses at each time. We
model this search among causal pathways decision-theoretically. Decision-theoretic
control usually results in significant improvements over unaided human expert judgments.
Furthermore, these improvements in performance are robust to substantial
errors in the assessed costs and probabilities. / Graduation date: 1995
Identifer | oai:union.ndltd.org:ORGSU/oai:ir.library.oregonstate.edu:1957/35300 |
Date | 06 June 1994 |
Creators | Yuan, Soe-Tsyr |
Contributors | D'Ambrosio, Bruce |
Source Sets | Oregon State University |
Language | en_US |
Detected Language | English |
Type | Thesis/Dissertation |
Page generated in 0.0017 seconds