Use this URL to cite or link to this record in EThOS: http://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.564046
Title: Probabilistic independence and its impact on the use of maximum entropy in casual networks
Author: Markham , Michael John
Awarding Body: University of Bradford
Current Institution: University of Bradford
Date of Award: 2005
Availability of Full Text:
Access from EThOS:
Abstract:
Causal Networks are suitable for some classes of probabilistic problems in Expert Systems but the methodology suffers from a variety of restrictions including the requirements for a specific set of information and a set of implicit independence relationships. These requirements must be satisfied to define the system uniquely and, as a consequence, other methodologies deserve consideration. Maximum Entropy is a candidate in this respect, it too has limitations, but this thesis examines ways of overcoming some of them. Entropy is maximised by using a method proposed by Tribus but this method only applies to linear constraints. The reliance of Causal Networks on implicit independencies gives rise to non-linear constraints and this opens up a field of research with two aspects:- The first aspect implies that the Method of Tribus be extended to include the handling of independence constraints. This would require the construction of a new algorithm. The second aspect calls for an investigation of the independence relationships in order to determine the minimum set of independencies that would have to be imposed on Maximum Entropy so that its solution becomes equivalent to that given by the Causal Networks methodology. Both aspects have been addressed in this thesis. An iterative algorithm to handle independence constraints has been developed, and techniques for finding a small set of independencies which will constrain the Maximum Entropy solution to match that given by Causal Networks have been devised. These techniques will also facilitate the estimation of missing information in incomplete cases.
Supervisor: Not available Sponsor: Not available
Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral
EThOS ID: uk.bl.ethos.564046  DOI: Not available
Share: