Metareasoning is a core idea in AI that captures the essence of being both human-like and intelligent. The idea is that much can be gained by thinking (reasoning) about one’s own thinking. In the context of search and planning, metareasoning involves making explicit decisions about computation steps, by comparing their `cost’ in computational resources, against the gain they can be expected to make towards advancing the search for a solution (or plan), and thus making better decisions. To apply metareasoning, a meta-level problem needs to be defined and solved with respect to a specific framework or algorithm. In some cases, these meta-level problems can also be very hard to solve (sometimes even harder than the original search problem). Yet, even a fast-to-compute approximation of meta-level problem solutions can yield good results and improve the algorithms to which they are applied.
This talk focuses on the development and evaluation of different metareasoning techniques, tailored for different problem settings, designed to improve a variety of search, planning and scheduling algorithms.