One of the most critical aspects of a truly intelligent system is the ability to learn, that is, to improve its own functionality by interacting with the environment and exploring it. In this paper, we argue that learning from exploring the environment should be the main goal in developing artificial intelligence. We also argue in favor of an integrated system-combining several stateof-the-art aspects of artificial intelligence, such as speech, vision, natural language, expert systems-as the experimental platform with which to approach this problem. We then describe the main features of a project of this type, MAIA, which is under development at I.R.S.T. The vision components of the system will be discussed in some detail, especially the navigation architecture of the indoor robot available to MAW. We will conclude outlining some initial learning problems that will be approached within the MAIA project, such as learning to recognize faces and learning to update the map of the Institute used for indoor navigation.
A New Definition of Artificial IntelligenceThe "Turing test" has represented for several years a definition of intelligence against which most workers in artificial intelligence (AI) have implicitly measured their own goals and achievements. It is an operational definition: If a computer behaves in a way indistinguishable from a human person, then it can be called intelligent. Recent criticisms of AI, for instance by Searle [l], can be summarized neatly by recognizing that they amount to questioning the validity of this definition of intelligence. Criticisms of this type are somewhat moot, since definitions are just definitions. It is, however, interesting to look for definitions of intelligence that are alternative or complementary to Turing's, not in order to claim that the computer can never be intelligent (as Searle and, more recently, Penrose [2] claimed) but in order to better capture the essence of the problem of creating artificial intelligence, at least as it is perceived today after 25 years of work in AI.Twenty-five years ago, intelligence was mainly reasoning, proving theorems, and playing chess. Today we realize how "intelligent" lower animals are and how complex are the problems that our senses routinely solve. We also realize how intractable is the problem of producing software and how much of it would be needed in order to replicate even just some of the simplest aspects of intelligence (think of the project by Lenat in Austin!). In this perspective, it seems natural to propose a somewhat different definition of intelligence. We suggest that this new definition should emphasize learning. Consider an artificial system such as a robot: We may define it as intelligent if it would be able to learn from exploring 0 1992 John Wiley & Sons, Inc.