Two papers authored by CHAI have been accepted for the Interantional Conference on Machine Learning 2019.
The first paper, On the Feasibility of Learning, Rather than Assuming, Human Biases for Reward Inference, is by Rohin Shah, Noah Gundotra, Pieter Abbeel, and Anca Dragan.
The second paper, Cognitive Model Priors for Predicting Human Decisions, is by David D. Bourgin, Joshua C. Peterson, Daniel Reichman, Stuart J. Russell, Thomas L. Griffiths
The thid paper is On the Utility of Learning about Humans for Human-AI Coordination by Micah Carroll, Rohin Shah, Mark K. Ho, Thomas L. Griffiths, Sanjit Seshia, Pieter Abbeel, and Anca Dragan. This paper has also been submitted to NeurIPS 2019.
The second paper has not been released for public review yet, but an abstract has been provided below:
Recently, deep reinforcement learning has been used to play Dota and Starcraft, using methods like self-play and population-based training, which create agents that are very good at coordinating with themselves. They “expect” their partners to be similar to them; they are unable to predict what human partners would do. This can be a big problem when trying to coordinate with humans, rather than playing against them. The authors demonstrate this with a simple environment that requires strong coordination based on the popular game Overcooked. They then show that agents specifically trained to play alongside humans perform much better than self-play or population-based training when paired with humans, particularly by adapting to any suboptimal human gameplay.