Date of Award
Doctor of Philosophy (PhD)
J. Bush Jones
The two problems in reconstructability analysis, abbreviated as RA, are referred to as the reconstructability problem and the identification problem. The former relates to the process of reconstructing a given system under a given criterion from the knowledge of its subsystems and, during this process, identifying those subsystems that are important in the reconstruction. The latter allows the identification of an unknown system from the knowledge of its subsystems. The advent of RA has intensified the research efforts on system studies. The objective of this research is to study the process of system reconstruction for general systems and apply it in the context of automated knowledge acquisition from databases. First, we describe basic concepts in reconstructability theory and machine learning. We then modify existing results in reconstructability theory for probabilistic and selection systems in order to generate better algorithms for determining the unbiased reconstruction and reconstruction families in the wake of new developments such as k-systems and the use of independent information. Further, we extend RA methodology for possibilistic systems using only partial information. An algorithm is proposed to compute the unbiased reconstruction, and the reconstruction families are identified as a set of max-min fuzzy relation equations. Furthermore, we define a new measure of the cognitive contents of a rule, referred to as the K-measure. Based on the K-measure, we introduce a new approach for automated knowledge acquisition from databases. Based on RA, the reconstructability approach to generalized rule induction from databases should work for most data covered by the framework of RA and k-systems. In particular, this approach is appropriate for expert-systems-like domains where the data is intrinsically nominal. Finally, we summarize our results and discuss the potentials for further research.
Trivedi, Sudhir Kumar, "Reconstructability Theory for General Systems and Its Application to Automated Rule Learning." (1993). LSU Historical Dissertations and Theses. 5678.