In late April I was too busy to join the thread "Circular definitions of intelligence". However, since some of you know that I proposed my working definition of intelligence before, I have no choice but to take Richard's challenge. ;-)
Before addressing Richard's (very good) points, let me summarize my opinions presented in http://www.cogsci.indiana.edu/pub/wang.intelligence.ps and http://nars.wang.googlepages.com/wang.AI_Definitions.pdf , especially for the people who don't want to read papers. First, I know that many people think this discussion is a waste of time. I agree that spending all the time arguing about definitions won't get AGI to anywhere, but the other extreme is equally bad. The recent discussions in this mailing list make me think that it is still necessary to spend some time on this issue, since the definition of intelligence one accepts directly determines one's research goal and criteria in evaluating other people's work. Nobody can do or even talk about AI or AGI without an idea about what it means. Though at the current time we cannot expect a perfect definition (we don't know that much yet), it doesn't mean any definition or vague notion are equally good. A good definition should be (1) clear, (2) simple, (3) instructive, and (4) close to the common usage of the term in everyday language. Since these requests often conflict with each other, our choice must be based on a balance among them, rather than on a single factor. Unlike in many other fields where the definition of the field doesn't matter too much, in AI it is the root of many other problems, since the only widely accepted sample of intelligence, human intelligence, can be specified and duplicated in several aspects or perspectives, and each of them lead the research to a different direction. Though all these directions are fruitful, they produce very different fruits, and cannot encompass one another (though partial overlaps exist). Based on the above general consideration, I define "intelligence" as "the ability to adapt and work with insufficient knowledge and resources", which requires the system to depend on finite computational capacity, to open to novel observations and tasks, to respond in real time, and to learn from experience. NARS is designed and implemented according to this working definition of intelligence. In the following I'll comment on Richard's opinions. On 4/26/07, Richard Loosemore <[EMAIL PROTECTED]> wrote:
I spent a good deal of effort, yesterday, trying to get you to "define intelligence in an abstract way that is not closely coupled to human intelligence" and yet, in the end, the only thing you could produce was a definition that either: a) Contained a term that had to be interpreted by an intelligence - so this was not an objective definition, it was circular,
Though "circular definition" should be rejected in general, this notion cannot be interpreted too widely. I'll say that defining "intelligence" by "mind", "cognition", "thinking", or "consciences" doesn't contribute much, but I don't mind people to use concepts like "goal" in their definitions (though I don't do that for other reasons), because "goal" is a much simpler and more clear concept than "intelligence", though like all human concepts, it has its own fuzziness and vagueness. Richard is right when saying that intelligence is required to recognize goal, but in that sense, all human concepts are created by human intelligence, rather than obtained from the objective world. Under that consideration, all meaningful definitions of intelligence will be judged as "circular". Even so, to define "intelligence" using "goal" is much less circular than using "intelligence" itself. Again, for our current question, no answer is perfect, but it doesn't mean all answers are equally bad (or equally good).
b) Was a definition of such broad scope that it did not even slightly coincide with the commonsense usage of the word "intelligent" ... for example, it allowed an algorithm that optimized ANYTHING WHATSOEVER to be have the word 'intelligent' attached to it,
Agree. If all computers are already intelligent, then we should continue to go with computer science, since the new label "AI" contribute nothing. According to my definition, a thermostat is not intelligence, and nor is an algorithm that provide "optimum" solutions by going through all possibilities and pick the best. To me, whether a system is intelligent is not determined by what practical problems it can solve at a given moment, but by how it solves problems --- by design or via learning. Among learning systems, to me the most important thing is not how complex the results are, but how realistic the situation is. For example, to me, a system assuming sufficient resources is not intelligent, no matter how great the result is. I don't think intelligence should be measured by problem-solving capabilities. For example, Windows XP is much more capable than Windows 3.1, though I don't think it is more intelligent --- to me, both of them have little intelligence. Yes, intelligence is a matter of degree, but it doesn't mean that any system will have a non-zero degree in this scale. BTW, I think it is too early to talk about numerical measurement of intelligence, though we can use the term qualitatively and comparatively.
c) Was couched in terms of a pure mathematical formalism (Hutter's), about which I cannot even *say* whether it coincides with the commonsense usage of the word "intelligent" because there is simply no basis for comparing this definition with anything in the real world -- as meaningless as defining a unicorn in terms of measure theory!
I think two issues are mixed here. To criticize the formalness of Hutter's work is not fair, because he makes its relation with computer system quite clear. It is true that he definition doesn't fully match the commonsense usage of the word, but no clear definition will --- we need a definition exactly because the commonsense usage of the word is too messy to guide our research. To criticize his assumption as "too far away from reality" is a different matter, which is also why I don't agree with Hutter and Legg. Formal systems can be built on different assumptions, some of which are closer to reality than some others. For example, it is possible to build a formal model with the assumption of infinite resources, and another one with the assumption of finite resources. We cannot say that they are equally unrealistic just because they are both formal.
In all other areas of science, a formal scientific definition often does extend the original (commonsense) meaning of a term - you cite the example of gravity, which originally only meant something that happened on the Earth. But one thing that a formal scientific definition NEVER does is to make a mockery of the original commonsense definition.
Again, it is a balance. I believe my definition capture the essence of intelligence in a deep level, though I acknowledge its difference on the surface level with the CURRENT commonsense usage of the word --- the commonsense usage of words do evolve with the progress of science.
I am eagerly awaiting any definition from you that does not fall into one of these traps. Instead, it seems to me, you give only assertions that such a definition exists, without actualy showing it. ********* Unless you or someone else comes up with a definition that does not fall into one of these traps, I am not going to waste any more time arguing the point. Consider that, folks, to be a challenge: to those who think there is such a definition, I await your reply. Richard Loosemore
So I've tried. I won't challenge people to find imperfectness in my definition (I know there are many), but do want to challenge people to propose better ones. I believe this is how this field can move forward --- not only by finding problems in the existing ideas, but also by suggesting better ones. Pei ----- This list is sponsored by AGIRI: http://www.agiri.org/email To unsubscribe or change your options, please go to: http://v2.listbox.com/member/?member_id=231415&user_secret=fabd7936