New Directions in Cooperative AI

In this talk, I will discuss some of the key challenges of collecting, leveraging, and interacting with humans in cooperative settings. Many prior works make extremely restrictive assumptions about human behaviors—assuming humans are rational decision makers; or assuming their suboptimality is only of a specific predefined type or is well-defined. In practice, these assumptions simply fail when learning from human data or when interacting with humans. This leads to most prior work to rely on machine-generated data or synthetic agents to demonstrate interesting behaviors. Here we would like to carefully analyze these assumptions, and study alternative approaches when there is a need for interaction and cooperation with real humans. This talk will consist of three parts: First we will talk about challenges of learning from humans and how we can tap into non-traditional human datasets. We will then talk about how to teach humans when different types of human suboptimalities and inductive biases are present, and finally we will discuss some open problems and future directions for online and interactive learning from humans especially when human data is variable.