Nonlinear Inverse Reinforcement Learning for Human Performance Enhancing Feedback

Many of the today’s most wicked problems are rooted in human behavior: whether it is distributing professional skills, managing chronic health conditions or responsibly sharing a common resource. Previous generations have dealt with such challenges through methods such as market forces, natural consequences and governmental policies. In this paper we propose a new, potential approach. We hypothesize that Inverse Reinforcement Learning can learn multifaceted, nonlinear reward functions which can drive predictable behavior change. To test this theory we create a stochastic, online control task and modify the task rewards according to IRL. A quasi-experimental design with a pre- and post-test, 4 treatment levels (plus a control group) and 400 participants per group (n≈2000) suggests that IRL rewards can cause behavior changes in a predictable manner.