Psychology-Informed Reinforcement Learning for Situated Virtual Coaching in Smoking Cessation

This thesis investigates how Reinforcement Learning (RL) can increase support effectiveness in virtual coach-based smoking cessation interventions. Such interventions have shown promise in helping people change behaviors such as smoking. However, personalizing the support they provide by accounting for people's current and future states might further increase their effectiveness. States thereby refer to people's relatively stable conditions at certain moments in time, capturing aspects such as motivation, knowledge, or the presence of personal reminders. After deriving general user needs for the support provided by a virtual coach-based smoking cessation intervention from a study with 671 daily smokers, we thus used RL to adapt the support to people's current and future states. Specifically, using data collected from three crowdsourcing studies with each more than 500 participants, we assessed the effectiveness of different RL model components in adapting 1) how people are persuaded, 2) what they are asked to do, and 3) who they are supported by. Our findings suggest that considering current and future states increases the effort smokers spend on smoking cessation activities and helps them build quitting-related competencies over time. Given that model components were derived from behavior change theories, this shows the potential of using psychology-informed RL to create smoking cessation support that is effective in the long run.