Modified: October 17, 2022
meta-reasoning
This page is from my personal notes, and has not been specifically reviewed for public consumption. It might be incomplete, wrong, outdated, or stupid. Caveat lector.Stuart Russell told the story of giving a talk on meta-reasoning at Stanford, with Don Knuth in the audience, where he opened with a slide that contained only the word "Algorithms" with a big red X over it. This is provocative, but I think it nicely illustrates the deep goal: we want systems that don't need to be programmed, that don't need to follow prespecified computational patterns (algorithms), because they can select online which computation will be most useful to do in that moment.Of course it may be useful to identify repeated subtasks and amortize the work of selecting computations that solve those tasks, thus recovering algorithm design. This seems key to building really flexible intelligence.
What kinds of tasks require metareasoning?
- Search in complex spaces. Planning the next move in a game of Go requires monte carlo tree search can be thought of as a metareasoning algorithm.