Decisions Based upon Multiple Values: The BVG Agent Architecture

摘要

Even a simple artificial agent will have to make decisions and consider their consequences. An agent can achieve this by evaluating situations along several dimensions. The BVG (Beliefs, Values and Goals) architecture is a means to build up such an evaluation machinery. The motivational part of the agent's mind is captured by the concept of value, able to capture the preferenes of the agent, which cannot be reduced to either beliefs or goals, or a combination of the two. We sketch a calculus for agent decision, and look into the problem of feedbacking the consequences of actions into the agent's decision mechanism, in order to adjust its behaviour for future occasions. Finally, we describe an experiment where these ideas are put to test and look into its results to draw conclusions about their general applicability.