Measuring Universal Intelligence in Agent-Based Systems Using the Anytime Intelligence Test

This paper aims to quantify and analyze the intelligence of artificial agent collectives. A universal metric is proposed and used to empirically measure intelligence for several different agent decision controllers. Accordingly, the effectiveness of various algorithms is evaluated on a per-agent basis over a selection of abstracted, canonical tasks of different algorithmic complexities. Results reflect the different settings over which cooperative multiagent systems can be significantly more intelligent per agent than others. We identify and discuss some of the factors influencing the collective performance of these systems.