On Automated Agents’ Rationality

Agents that interact with humans are known to benefit from integrating behavioral science and exploiting the fact that humans are irrational. Therefore, when designing agents for interacting with automated agents, it is crucial to know whether the other agents are acting irrationally and if so to what extent. However, little is known about whether irrationality is found in automated agent design. Do automated agents suffer from irrationality? If so, is it similar in nature and extent to human irrationality? How do agents act in domains where human irrationality is motivated by emotion? This is the first time that extensive experimental evaluation was performed in order to resolve these questions. We evaluated agent rationality (for non-expert agents) in several environments and compared agent actions to human actions. We found that automated agents suffer from the same irrationality that humans display, although to a lesser degree.