| Engineering properties of soils and rocks are evaluated using tests which, at best, approximate stress and drainage boundary conditions in situ (and hence may not accurately reflect soil behavior under field conditions), or by means of correlation with index properties. Once one or more of these uncertain indications of a property are available, the engineer must form a judgment in order to select the appropriate value or range of values to use in analysis. This dissertation describes the development of a framework for representing those judgments in artificial intelligence using Dempster-Shafer evidence theory, an extension of classical probability theory. This approach treats not the probability that a conclusion is correct, but the probability that the evidence indicates that the conclusion is correct. This view of probability allows one to quantify ignorance or doubts about the means of predicting the engineering properties, and to combine test data with subjective beliefs regarding what the data indicate about soil behavior. Support from different pieces of evidence is synthesized into a single support distribution by means of Dempster's rule and conflict among the pieces of evidence is resolved probabilistically. Since Dempster-Shafer theory had previously only treated discrete outcomes (e.g. yes or no), it was necessary to derive a modification for continuous variables. That modification was verified by comparison with judgments by experienced engineers. The modified Dempster-Shafer theory is also a logical framework for the interpretation of testing and correlation results outside of expert systems as well, particularly when the data are few. |
