Optimal agents

Abstract

Artificial intelligence can be seen as the study of agents which achieve what we want, an agent being an interactive system with an input/output interface. This can be mathematically modelled through decision theory, using utility functions to capture what we want, and with the expected utility of an agent measuring how well it achieves that. Optimal agents, those that maximally achieve what we want, have maximal expected utility. We describe this theory of optimal agents. We detail how these optimal agents behave, giving an explicit formula for their actions. We discuss applications of this theory, in particular Marcus Hutter’s AIXI [Hut04]. Finally, we suggest possible directions for future research extending Hutter’s work on universal AI.