Analogy and metareasoning: Cognitive strategies for robot learning

Abstract

Robots share a conundrum central to all intelligence. Like humans, robots not only must address novel situations but also must start from what they already know: how, then, can any robot deal with novelty? In this chapter, we examine two cognitive strategies for addressing novelty: analogy and metareasoning. Analogy addresses new problems in a manner similar to a familiar problem; the familiar problem is part of the context of addressing the new problem. We show how a cognitive robot may use analogy to learn from a small number of initial demonstrations. However, analogical reasoning and reasoning more generally do not necessarily guarantee success in an open, dynamic world. This brings metareasoning into play. We show how a cognitive robot may use metareasoning to recover and learn from failure; here, failure forms part of the context for metareasoning. In the other direction, we discuss how these experiments in robot learning inform the development of cognitive theories of analogy and metareasoning.