Decision making by agent committee

Abstract

Computer decision making systems are aids that are becoming used in an increasing number and variety of applications. These systems allow people to interface with sources of information far too large and complicated to process themselves. As these systems become used on more advanced and complicated tasks, they will need to become more intelligent. One step towards this is the creation of a multi-agent decision making system that uses behavior inspired by the interactions of groups of people to allow a set of agents with humanistic characteristics to interact with one another. I implemented a program, AgentCommittee, which incorporates such behavior. AgentCommittee uses a set of characteristics that include extroversion, fatigue, resistance, confidence, and competitiveness in a series of one-on-one interactions to arrive at a group decision. The goal of AgentCommittee is not to find the best or optimal answer, but to produce a variety of good answers. I tested this program on a set of data from Compaq's EachMovie database and found that AgentCommittee was 20% more successful at finding relationships between the genre of a movie and a user's opinion of that movie than a random output generator.