Human Models for Planning Behavioral Interventions with Reinforcement Learning

Abstract

Many important areas of behavior change, such as wellness or education, are frictionful; they require individuals to expend effort over a long period of time with little immediate gratification. Because of this, humans often act sub-optimally concerning their stated long-term goal. Here, an artificial intelligence (AI) agent can provide personalized behavioral interventions to correct human policies. The AI must personalize rapidly (before the individual has a chance to disengage) and interpretably, to aid our scientific understanding of the behavioral interventions. This work focuses on crafting small, interpretable models of the human that capture the mechanism behind the human agent's sub-optimal policies. These human models provide the AI with enough inductive bias to quickly learn intervention policies for each individual it encounters.