This paper provides a new framework for viewing explanations as arguments supporting different model predictions, rather than proofs that such predictions are correct. We provide user studies to show that in the case where we want humans to learn from AI explanations while making the final decision, providing dissenting explanations generated from models making opposing predictions allows them to significantly reduce overreliance on the original model, without significantly reducing accuracy. I worked on this project with PhD student Judy Shen, in Omer Reingold's algorithmic fairness lab. It was presented at the AI & HRI workshop at the International Conference on Machine Learning in July 2023.

FirePOMDP

In this paper, we model the spread of wildfires as a partially-observable Markov decision process, where a wildfire starts in some cells of a grid world and then spreads as a random variable according to factors such as wind. We assume that firefighters have limited resources in terms of putting out fires, and that some cells, representing residential areas, are more important to protect than others, and then use POMDP solver methods to learn the optimal way to fight the fire and limit spread. This was a final project for CS 238: Decision Making under Uncertainty.