Adaptive Control: From Feedback to Debugging

Abstract

I’m going to talk about control systems of the sort that ray colleagues work on, in this field called Artificial Intelligence. From this perspective, I tend to think about control in terms of knowledge. When you talk about tuning or hill climbing, the system is changing its state, and you can say that it knows different things after it’s got whatever it’s controlling under control. If you’re trying to have adaptive control, then the knowledge of the system keeps changing; and if you have expert control, which is the sort of thing one looks for in humans, then as the situation changes you learn general strategies for bringing about desired changes, so that long-term learning certainly can be seen.