Uncertainty management in rule based systems with applications to computer vision
Advait Mogre · 1990
This thesis introduces a unified approach to uncertainty management in rule based systems in terms of uncertainty modelling and propagation, context incorporation and conflict resolution. The application emphasized is that of Automatic Target Recognition (ATR) using a rule based approach. The first part of this thesis concerns the construction and implementation of a measure of certainty in a given proposition. This measure accepts objective evidence, in the form of feature information and subjective evidence or "expert opinion" for a given proposition: "Unknown (object, U) belongs to class 'A' given evidence x". The measure then combines the evidences, from both sources if available, to generate either a certainty value or a truth term (Chapter 4), depending on the model used to express confidence in the given proposition. This approach, using belief measures to express confidence in the association of a given class with the objective evidence, has been applied to sequences of images acquired by a forward looking infrared sensor (FLIR). The results show an improvement in the classification performance with respect to standard belief maintenance. The next portion of the research deals with the inclusion of contextual information. This involves constructing functions to combine contextual information and to integrate the resultant information into the confidence generation process. Results of simulations to study the effect of context on both numeric as well as linguistic confidences are presented. The effect of context is achieved by incorporating peripheral context rules that accept and quantify context information applicable to specific rules in the main rule base. The advantage of this approach is that the main rule base need not be modified each time additional context information has is acquired. Instead, the additional information is absorbed into the rule base by constructing additional context rules. The final aspect of this research concentrates on the problem of conflict resolution in combining evidence form different sources. First, the compensation technique used by Shafer, called "discounting" is discussed. Next, a measure of disparity within a group of evidence sources, for the same given proposition is presented. The measure of disparity enumerates the "differences of opinion" amongst the group, which is then used to generate a discount factor for that group. This technique is then successfully deployed to the ATR problem to improve classification. This approach is later extended to the case of disparity in linguistic truth qualification. Simulation results for this case are then presented. When appropriately utilized, each of these techniques introduces additional improvement in the performance of the decision making process for the ATR problem using a rule-based approach. However, they can be applied in a similar fashion to other problem domains as well.