A Value-Driven System for Autonomous Information Gathering

by Grass, J. and Zilberstein, S.

Grass and Zilberstein introduce a framework for gathering information, by repeatedly selecting queries with the highest marginal value. In contrast to coverage (precision/recall) based approaches value-driven systems maximize the user's value.

Value-driven information gathering (VDIG) complements current work in search on the Internet, as it focuses on the query selection problem using response time and cost as opposed to the coverage problem.

The authors identify five characteristics required for applying value-driven information gathering to a particular domain.

• a decision model is available
• the model is applicable to determine the value of information
• there are multiple information sources with different cost and response times
• the user operates with limited resources (time and money)
• information sources return information usable for the decision model

VDIG evaluates information based on

• evaluation of the decision model,
• the available resource, and
• the user's preferences (required to map time to utility)

Another interesting point is, that the authors learn the responsiveness of information sources by just querying these information sources for a number of time (an extension to this approach would be adjusting the data based on the last queries performed).

The cost structure imposed by information sources has a major influence on the behaviour of VLDIs. Costly information leads to more selective, lengthy processes with fewer queries.