Five Dimensions of Reasoning in the Wild

Reasoning does not work well when done in isolation from its significance, both to the needs and interests of an agent and with respect to the wider world. Moreover, those issues may best be handled with a new sort of data structure that goes beyond the knowledge base and incorporates aspects of perceptual knowledge and even more, in which a kind of anticipatory action may be key. Out of the Ivory Tower Reasoning is one of the oldest topics in artificial intelligence (AI). And it has made lots of progress, in the form of commonsense reasoning (CSR), planning, automated theorem-proving, and more. But I suspect it has hit a barrier that must be surmounted if we are to approach anything like human-level inference. Here I give evidence for such a barrier, and ideas about dealing with it, loosely based on evidence from human behavior. In rough synopsis, reasoning does not work well when done in isolation from its broader significance, both for the needs and interests of an agent and for the wider world. Moreover, those issues may best be handled with a new sort of data structure that goes beyond the knowledge base (KB) and incorporates aspects of perceptual knowledge and even more, in which a kind of anticipatory action many be key. I suspect this has ties with recent calls to “put the Science” back in AI (Levesque 2013, Langley 2012). For what I am arguing, in some sense, is that reasoning should be regarded as “in the wild” as events unfold rather than confined to management of an isolated KB; and that this speaks to an agent interacting with the world, rather than a puzzle in abstract inference (yet I will also argue that even “pure” reasoning as in mathematics hugely benefits from many connections with the world). And finally, we then will end up studying the nature of world-embedded cognitive agents, humans included. But this is very broadbrushed and general, whereas my main point is a technical suggestion about reasoning informed by meaning, especially meaning concerning experience and action. This is a slightly modified version of a paper that appeared in AAAI2016. One quick example at the outset: The Wason Selection Task (Wason 1968) shows that human inference is strongly aided when the details of the task at hand have real meaning that the subjects can relate to in terms of things that matter to them, helping keep attention on what is relevant; and this holds even when the task in the abstract is a matter of so-called pure logic. While this could be seen as a defect in human reasoning, something computers would never trip up on, I think it points in the opposite direction: inference without broader meaning is not worth much, and not worth being good at. I will illustrate my main points with a series of examples based on the activities of proving, planning, and understanding.