Reinforcement Learning Interventions on Boundedly Rational Human Agents in Frictionful Tasks

Area of Science:

• Artificial Intelligence
• Behavioral Science
• Reinforcement Learning

Background:

• Many crucial behavior changes are difficult, requiring sustained effort without immediate rewards.
• Artificial intelligence (AI) offers potential for personalized interventions to support goal adherence.
• AI agents need to personalize interventions rapidly and interpretably for effective behavioral support.

Purpose of the Study:

• Introduce Behavior Model Reinforcement Learning (BMRL) for AI-driven behavioral interventions.
• Model human decision-making as a planning agent within a Markov Decision Process (MDP).
• Attribute suboptimal human policies to maladapted MDP parameters.

Main Methods:

• Developed a framework where AI intervenes on the MDP parameters of a boundedly rational human agent.
• Proposed tractable human models capturing behaviors in effortful tasks.
• Introduced MDP equivalence specific to BMRL.

Main Results:

• AI planning with proposed human models leads to effective policies.
• Demonstrated theoretical and empirical support for the BMRL framework.
• Showed AI can help individuals stick to goals in complex, real-world scenarios.

Conclusions:

• BMRL provides a novel approach for AI to support long-term behavioral changes.
• Interpretable AI interventions can enhance understanding of behavioral dynamics.
• The framework effectively addresses challenges in frictionful tasks requiring sustained effort.