On the intelligence of interacting autonomous robots and virtual agents

Abstract

[EN] This work explains some aspects why it is hard to pinpoint what intelligence is and more specifically, how to assess the intelligence of AI. It motivates a setup that is designed to foster the investigation of this question using reinforcement learning agents as complex AI systems. Such a setup can be used in an attempt to sidestep theoretical considerations on the cognitive power of Machine Learning algorithms. Instead, an example is given how the well-established experimental testing of intelligence in animals could be translated to the described AI system. While the published work-in-progress state of the implementation allows similar experiments of multiple interacting virtual robots to be conducted and a theoretical outline for future tests is sketched, a lot of further research will be required before a robot can demonstrably recognize itself in a mirror.