Commonsense Physics: A Review
Kenneth D. Forbus
Qualitative Reasoning Group, Department of Computer Science, University of Illinois, Urbana
Ann . Rev . Comput. Sci. 1988 .3, 197-232
Copyright ©1988 by Annual Reviews Inc (All rights reserved)
1 . INTRODUCTION
Understanding commonsense reasoning is a central problem of Artificial Intelligence. Without a broad codification ofhuman knowledge, and techniques for reasoning with such knowledge, our programs are doomed to remain confined to specialized areas. While expert systems have sometimes been strikingly successful in narrow, carefully defined domains, they remain brittle and hard to maintain . Natural language „front ends“ are successful only when the domain of discourse is strictly limited. Robots cannot predict indirect consequences of their actions (e.g. that leaving a tool outside may cause it to rust). Really smart programs, especially those that must interact frequently with human beings, must share our common knowledge and assumptions .
What is commonsense reasoning? Sometimes it is defined only indirectly, by contrast with „expert reasoning .“ Some identify commonsense reasoning with default or nonmonotonic reasoning. Neither definition seems appropriate. Psychological studies ofexpert reasoning indicate that it relies heavily on our mental models, our commonsense theories of the domain of expertise (Gentner & Stevens 1983; de Kleer & Brown 1984). While experts certainly know more than novices about their domain, typically this additional knowledge is highly interconnected with the knowledge that both share. This suggests that we need to say more directly what that knowledge is. And while commonsense reasoning often involves defaults and nonmonotonic reasoning, it is hard to find areas of expertise that do not also involve defaults and nonmonotonicity . Hence such reasoning cannot be a defining property of commonsense.
- 5. Juli 2013