Value-driven information gathering

• Main Author: Grass, Joshua William
• Language: ENG
• Published: ScholarWorks@UMass Amherst 2002
• Subjects: Computer science
• Online Access: https://scholarworks.umass.edu/dissertations/AAI3039360

This dissertation addresses the problem of autonomous information gathering from a large distributed network of information sources. Information gathering is viewed as a component of a decision support system, which uses a set of rules and a set of information sources to recommend an action. This recommendation includes a prediction of the utility for selecting this action and the level of confidence that the system has in the decision. A decision support system has two primary tasks when making a well-informed, well-reasoned decision. The first task is to gather information about the state of the world that is relevant to making the decision; and the second task is to use this information and a set of rules to evaluate a set of potential actions and make a recommendation. A large number of information gathering systems have been developed in recent years that use the Internet as their primary source of information. However, the overwhelming amount of information available on the Internet has created a new problem for information gathering systems: it is no longer feasible to query and process all of the available relevant information. Next-generation information gathering systems must account for the resources required to query and process the information sources used by the system. To address this problem, this dissertation develops a decision-theoretic framework for information gathering that is sensitive to several characteristics of information sources. These characteristics include the value of acquiring a piece of information with respect to the specific user's decision model, the strength of the evidence returned by the information source, the immediate cost of querying the information source, and the expectation of when and if the query will return information. The comprehensive value of a query, which is an extension of the decision-theoretic notion of the value of perfect information (VOI), is calculated using these characteristics. Much like the VOI, the value of a query is based on the notion of determining the expected increase in the overall expected utility of the decision as a result of issuing the query. However, unlike the VOI, the value of a query reflects the fact that information gathering is not instantaneous and may have associated costs. There are three main contributions made by this dissertation. The first contribution is the development of a formal framework for query planning with limited information gathering resources that is driven by the user's decision model (an influence diagram). The second contribution is implementing this framework as an expandable system for creating autonomous information gathering agents. The third contribution is demonstrating how value-driven query planning yields improved information gathering strategies that return high-quality decisions while using substantially fewer resources. As the number of information sources available to autonomous information gathering systems grows, the role of reasoning about both the cost and benefits of querying any potential information source becomes increasingly important.