Modeling cognition as querying a database of labeled beliefs
Annerieke Heuvelink · TNO Repository · 2006
One of the most important conditions for an agent to ensure that it selects the right plans for fulfilling its goals, is that its knowledge about its own status and that of the world is consistent, and as correct as possible. The knowledge of an agent is referred to as its belief set.The common technique within AI to deal with inconsistencies in belief sets is to throw away the beliefs that cause it. In doing so, the agent forgets about what it believed to be true before, which is not very human-like. Furthermore, often just one way to resolve an inconsistency is proposed, while cognitive science has shown that the way in which humans deal with inconsistencies is influenced by belief properties, like the nature and the source of the beliefs. (e.g., Dieussaert et al., 2000; Mercier & der Henst, 2005). Moreover, humans often do not end up disbelieving statements, but rather revise them. For example, they extend a conditional rule with an extra condition or an exception clause (Walsh & Sloman, 2004).Although humans can reason about what they believe to be true and why, it has been found on multiple occasions that they do not always do so in a rational way. Humans,especially under pressure, often get biased. For example, decision makers tend to be biased by the order in which pieces of information arrive (e.g., Anderson, 1981).This paper introduces a belief modeling approach for agents that enables a more realistic handling of (inconsistent) beliefs than is currently standard in AI.Beliefs are not thrown away, but get labeled with a time stamp, their source and a certainty level. These properties enable the modeling of a great variety of realistic reasoning strategies, ranging from rational ones that generate optimaloutcomes to false ones that might yield biased outcomes. The resulting agent, capable of generating correct as well as biased outcomes, can be used to support training.