Encoding Intelligent Agents for Uncertain, Unknown, and Dynamic Tasks: From Programming to Interactive Artificial Learning

In this position paper, we analyze ways that a human can best be involved in interactive artificial learning against a back-drop of traditional AI programming and conventional artifi-cial learning. Our primary claim is that interactive artificial learning can produce a higher return on human investment than conventional methods, meaning that performance of the agent exceeds performance of traditional agents at a lower cost to the human. This claim is clarified by identifying met-rics that govern the effectiveness of interactive artificial learn-ing. We then present a roadmap for achieving this claim, identifying ways in which interactive artificial learning can be used to improve each stage of training an artificial agent: configuring, planning, acting, observing, and updating. We conclude by presenting a case study that contrasts program-ming using conventional artificial learning to programming using interactive artificial learning