A Self-Help Guide For Autonomous Systems

plan or hope; deviations from the ideal are a fact of life. As Heraclitus noted almost 2500 years ago, “only change endures.” Though Heraclitus was silent on whether the inevitable change will be good or bad, Murphy’s Law tells us to expect the worst. Given the pervasive and often perverse nature of the unexpected, it is not surprising that humans are adept at dealing with it—we would not have survived and flourished otherwise. This facility with the unexpected may account for our annoyance when our automated systems break in situations that we could deal with easily. It is then we need to remember that such failures account for much of the progress in AI, since—as this special issue attests—these annoying surprises goad us into insights and thus to improvements in the underlying algorithms. But one can also view this phenomenon in another light. Instead of our having to rebuild the algorithms when—in an unanticipated situation—they make “stupid” mistakes that we could easily have dealt with, why not build them to do what we do: deal with the mistakes on their own? This is a tall order, and not a new one: it requires addressing the so-called brittleness problem (see Brachman [2006]). In this article, we present our own approach to this problem. First we describe the conceptual focus—which we call the metacognitive loop (MCL)—that is guiding our work; then we present a number of domain-specific implementations of MCL, which we think provide significant evidence that MCL is a general-purpose methodology for building flexible (nonbrittle) programs in specific domains; and finally we discuss our current efforts toward a domain-independent implementation of MCL.