Goal

Goal pursuit and goal attainment are at the core of individual, and by extension societal, success and flourishing. Existing research, however, shows that many of the goals people set are not attained, and that self-control is often ineffective for resisting temptations. By bringing together researchers from psychology, behavioural science, and computer science, it should be possible to develop a toolbox of effective goal pursuit and self-regulation strategies that can be automatically customized based on the individual, the goal, and the situation. One proposed project could be to first create a toolbox of empirically-supported strategies to improve goal pursuit, and test these strategies in a mega-experiment (similar to https://bcfg.wharton.upenn.edu/ourapproach/). Then, using that data and machine learning, develop a mobile application that actually provides users with a prompt to help self-control based on characteristics of the individual, the goal, and the situation. Reinforcement learning could be used so that the app can 'learn' to customize proposed intervention based on what is effective for the user, akin to personalized medicine.
By learning to set better goals and improving goal pursuit, individuals can improve in a variety of domains, including health, education, personal savings, and relationships, to name a few.

Opportunity

Previous research has tested interventions to improve goal pursuit. However, they either focused on specific goals (e.g., exercise, quitting smoking), or used a one-size-fits-all approach. Additionally, although many mobile applications to improve goal pursuit exist, they are not empirically supported. The possibility of using machine learning together with large amounts of data derived from a within-person mega-experiment combines new directions in intervention studies with approaches derived from personalized medicine. Designing an app that can 'learn' what is most likely to work for any given goal and situation is a completely new approach to personalized intervention to improve self-regulation.

Roadblocks

The sheer scope of the endeavor is the largest challenge, and required sufficient resources. More specific challenges to overcome includes: (1) reviewing all interventions from different disciplines that could be effective and adapting them for idiosyncratic goal pursuit (2) Technological challenges inherent in using machine learning (e.g.,, understanding users' indicated idiosyncratic goals), and refining the algorithm to effectively select the most appropriate intervention to suggest. (3) Participant recruitment and retention – in the first phases, this research would require participants who would use the app regularly (multiple times/day) for multiple weeks, to obtain sufficient within-person data to apply machine learning.

Breakthroughs Needed

Resources provided by the foundation would help recruit sufficient personnel and assemble a multidisciplinary team of researchers capable of tackling such a complex project. Bringing together social and health psychologists, behaviour scientists, and behaviour economists with experience in goal pursuit and behaviour change (across multiple types of behaviours) would be a first step to ensure that a broad range of possible interventions are obtained. Working closely with computer scientists and developers would help to overcome potential technological challenges. Designing the app to be engaging and optimize user experience, in addition to sufficient participant compensation, will help with recruitment and retention. Overall, the problem is not so much that there are not sufficient possible interventions, but that existing interventions are each examine separately, and are different for different types of goals/behaviours. The main breakthrough needed is for researchers to come together and examine relative efficacy of these various interventions, and especially figure out how to help individuals use them flexibly (i.e., use the best one for the person, goal and situation).

Key Indicators of Success

Given the high failure rates both of momentary self-regulation and of general goal-pursuit typically found in research, increasing success rates of both (as indicated by the success rates of app users) would be a key indicator of the project's success. In the short term, the ability to determine which interventions are most effective for whom and when, and to apply an algorithm that can figure it out across idiosyncratic goals (i.e., in developing an effective app) would constitute success. In the longer-term, another key indicator would be broad uptake (if the app is found effective, and are broadly rolled out).

Additional Information

Possible collaborators: Angela Duckworth; Wilhelm Hofmann; David Poole; Dror Ben-Zeev;

Technical summary:
Prior research has designed and tested interventions to promote behaviour change, particularly in the area of health behaviours. Many such interventions can be considered 'nudges' – a small change in the environment designed to encourage the person to act in a different way (Thaler & Sunstein, 2008). For example, a recent meta-analysis finds that that randomized controlled trials using mobile health tools had a small effect on physical activity. However, authors note that few mobile health (mHealth) interventions used goal setting, and most were based on short message service (text message) delivery. In another review, researchers highlighted that few mHealth tools employ theoretically driven behaviour change principles (Conroy et al., 2014). As such, there is a clear need for easily-implemented brief interventions that are developed based on theory.
In a proposed project, we will (1) examine and compare the effectiveness of multiple theoretically-derived interventions (delivered over a custom-made mobile application) to improve goal attainment of idiosyncratic (user-reported) personal goals; (2) using a machine learning model, determine whether such nudges could be 'matched' to goal and personal characteristics to maximize effectiveness; and (3) using data obtained in objective 2, refine the mobile application using reinforcement learning to continuously improve the matching quality. The final mobile application should be capable of using characteristics of the person and goal to provide customized prompts that become further adapted based on ongoing data collection (as the person continues to use the app).
Data will be collected from participants over multiple weeks for many personal goals; a different intervention will first be evaluated for each goal, with the participant providing feedback on effectiveness. To determine which nudge(s) work better based on personality and goal characteristics, the learning module in the machine learning approach, called the learning agent, evaluates whether a nudge choice is good or bad for an individual based on previous feedback instances (e.g., whether the person achieved the goal based on a specific nudge). Each individual's pre-collected data (including gender, to conduct gender-based analyses) and the type of goal (e.g., health, finance, etc.) forms a state, and each nudge is an action. The feedback received at the end of the week (achieved the goal or did not achieve the goal) is the reward for taking the action at this state. By aggregating rewards over different actions at different states, the learning agent can estimate with high probability which action (nudge) is more effective at which state. The app can then be further refined to use a reinforcement learning approach to continue integrating new information on goal attainment provided by the user to refine nudges presented in the future. That is, this app will understand the goal, and start by presenting nudges that are statistically most likely to work for that person and then adapt its suggestions based on what actually works (and does not work) for that person. The effectiveness of this customized app will be evaluated.