Learning Intended Cost Functions: Extracting all the right information from all the right places

Anca Dragan
University of California, Berkeley (UC Berkeley)

Optimal control work tends to focus on how to optimize a specified cost function, but costs that lead to the desired behavior consistently are not so easy to specify. Rather than optimizing specified costs, which is already hard, machines actually have the much harder job of optimizing intended costs. While the specified cost does not have as much information as we make our machines pretend, the good news is that humans constantly leak information about what the machine should optimize. In this talk, we will explore how to read the right amount of information from different types of human behavior -- and even the lack thereof.

Back to Intersections between Control, Learning and Optimization