The Theory, Implementation, and Practice of Evidential Reasoning
by Lowrance, John D. and Strat, Thomas M. and Wesley, Leonard P. and Garvey, Thomas D. and Ruspini, Enrique H. and Wilkins, David E.
Technical Note SRI Contra
Institution: AI Center, SRI International
Address: 333 Ravenswood Ave., Menlo Park, CA 94025
Evidential Reasoning (ER) is a body of techniques for automated reasoning from evidence that is based upon the mathematics of Dempster-Shafer belief functions. The emphasis of this project was twofold: to broaden and solidify the theoretical basis of ER, and to facilitate the transfer of the intellectual technology embodied in ER. As part of our theoretical effort, we established a sound semantics for Dempster-Shafer belief functions, deriving the Dempster-Shafer axioms based upon epistemic logic; we derived all ER operations from these same Dempster-Shafer axioms; we established ER as a generalization of both logical and Bayesian probabilistic reasoning; we identified the conditions under which the computational complexity of ER belief networks can be reduced; we developed a theoretically justified means of propagating all information throughout an evidential analysis, determining the global impact on all probabilistically dependent random variables; we incorporated decision theoretic concepts into ER, including sensitivity analysis and decision trees. To facilitate the transfer of this technology, we developed intuitive graphical structures for representing and manipulating evidential knowledge, thereby, substantially reducing the time required to compile and organize the knowledge for a new application domain, and we developed a set of canonical ER examples covering a range of application domains, including underwater vehicle tracking, antiair threat identification, medical diagnosis, and robot vehicle navigation.
|Garvey, Thomas D||Associate Director|
|Lowrance, John D||Program Director|
|Ruspini, Enrique H||Alumnus|
|Wilkins, David E||Senior Computer Scientist|