# Action Centered Contextual Bandits

**Authors:** Kristjan Greenewald, Ambuj Tewari, Predrag Klasnja, Susan Murphy  
**arXiv:** [1711.03596](https://arxiv.org/abs/1711.03596)  
**Date:** 2017-11-09  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/1711.03596v1/

## Key points

- The paper explores the use of contextual bandits in mobile health applications
- Contextual bandits provide a middle ground between simple multi-armed bandit approaches and complex reinforcement learning methods
- They have been successful in web applications due to their interpretability and ease of implementation, as well as strong performance guarantees when the linear model assumption holds true
- However, this assumption is not feasible in emerging mobile health applications
- The authors propose an extension of the linear model for contextual bandits that consists of two parts: baseline reward and treatment effect
- The theory presented in the paper is supported by experiments conducted on data gathered from a recent mobile health study
- This paper contributes to advancing contextual bandit algorithms for mobile health applications by accommodating nonlinearity in baseline modeling while preserving strong performance guarantees similar to those offered by linear models.

## Summary

The paper titled "Action Centered Contextual Bandits" by Kristjan Greenewald, Ambuj Tewari, Predrag Klasnja, and Susan Murphy explores the use of contextual bandits in mobile health applications. Contextual bandits provide a middle ground between simple multi-armed bandit approaches and complex reinforcement learning methods. They have been successful in web applications due to their interpretability and ease of implementation as well as strong performance guarantees when the linear model assumption holds true. However, in emerging mobile health applications this assumption is not feasible. To address this limitation, the authors propose an extension of the linear model for contextual bandits that consists of two parts: baseline reward and treatment effect. While the former can be complex, the latter remains simple yet allows for nonlinearity in baseline modeling while maintaining simplicity and interpretability. The theory presented in the paper is supported by experiments conducted on data gathered from a recent mobile health study. This paper contributes to advancing contextual bandit algorithms for mobile health applications by accommodating nonlinearity in baseline modeling while preserving strong performance guarantees similar to those offered by linear models.