Using Reinforcement Learning to Personalize Daily Step Goals for a Collaborative Dialogue with a Virtual Coach

Using Reinforcement Learning to Personalize Daily Step Goals for a Collaborative Dialogue with a Virtual Coach

Dierikx, Martin (TU Delft Electrical Engineering, Mathematics and Computer Science)

Brinkman, W.P. (mentor)
Albers, N. (mentor)
Pera, M.S. (graduation committee)

Delft University of Technology

Computer Science

2023-10-06

To reduce morbidity and mortality caused by multiple chronic conditions, the number of steps people take each day should be gradually increased. For this, a recommended step goal can be created that is based on an individual's previous walking behaviour. However, for a person, the achievability of this recommended goal can change daily because of that person's state, such as their mood or self-motivation. It could be, for example, that if someone's self-motivation is low, proposing a lower goal than the recommended one, increases their self-motivation and allows them to achieve the recommended goal the next day. Therefore, we investigated the use of a person's state to personalize daily step goal proposals. To do so, we designed and implemented a virtual coach, named Steph, to propose daily step goals to people during an observational study. We used people's states collected in that study to train a reinforcement learning model to optimally personalize the step goal proposals. Based on simulations of our model, we found that people in high states (e.g. who were very motivated and had a positive mood) were more likely to achieve their recommended goals, while people in low states (e.g. who were not motivated and had a negative mood) were less likely to achieve their goals. We also found that proposing higher goals to people in certain states was better than for people in other states. This was because, for some people, a higher goal improved their state while for others, it worsened it. This suggests that personalizing people's step goal proposals optimally could change people's states to where they are more likely to achieve their recommended step goals. So, this thesis provides a model for personalizing daily step goal proposals which can be used as part of behaviour change support systems. It can also serve as a basis for different approaches to predict and change people's walking behaviour to make them more active and less susceptible to chronic diseases.

Reinforcement Learning
Virtual Coach
Goal-setting
Physical Activity
Personalization
Collaboration

http://resolver.tudelft.nl/uuid:4f2c12de-9b9f-4e3f-ad3a-902947d693bb

Student theses

master thesis

© 2023 Martin Dierikx