peter@dai.ed.ac.uk Dept. of Artificial Intelligence, Univ. of Edinburgh.
~Why did you choose a career in AI research / development? For the interest, of course. AI is very highly interdisciplinary, with links to linguistics, psychology, philosophy, mathematics, computing and many more areas. There are a lot of deep questions still awaiting answers within AI; but there's also the chance to get involved in very worthwhile practical applications of the technology. You won't ever get paid a fortune but you will have a lot of fun.
~What is the AI system you are researching/developing designed for? This question supposes that there is just one. Most recently I've been involved in research on timetabling and scheduling by evolutionary means. We can get excellent results on a broad range of large problems, and this technology is now being put into practical use. But in the longer term the evolutionary computing group in Edinburgh are involved in many different applications of genetic algorithms, genetic programming and so on, and this all contributes to a better understanding of the potential of evolutionary computing.
~What approaches are you using in your research / development ? For specific projects, typically we study performance on classes of artificially-generated problems, and on real-world problems, and on artificially-generated variants of real-world problems. This means we routinely do hundreds of thousands of runs and then analyse performance statistics, anomalous cases, etc. But it's also important to study individual runs in great detail.
In the longer term, as the previous answer suggests, it's more a matter of gathering experience by trying many different things over years. Some questions are scientific: does an idea work? why or why not? how well could it ever work? what are its weaknesses? Other questions are technological: can it be made useful? can it be done cheaply and reliably enough? and so on. And then there are the very fundamental issues about how living brains work so well, which would seem to require some completely new ideas and som new approaches to studying and even just talking about enormously complicated dynamical systems.
~What do you see as some fundemental ways that AI in general will impact people's lives in the future? More of the same -- there are already large numbers of expert systems in practical use, and many neural and fuzzy systems. There are significant numbers of NL systems, planners, autonomous robot systems also in practical use. It would be pretty useless to speculate in detail; there will be ever more `smart devices'.
However, I would hope that there will be very large knowledge bases in the future, somewhat along the lines being pioneered by the CYC project. The CYC project does itself no great favours by displaying examples in which the question "Show me somebody wet" leads, albeit correctly, to a picture of a very sweaty athlete; the average person in the street would regard this as almost trivial but perhaps slightly entertaining, having no idea of the underlying technical issues and no way to judge whether this example is representative of its ability to handle many kinds of question.
A better scenario might be the following true-life one. Back in the 30s when gas pipelines were being constructed across hundreds of miles of desert in central and southern parts of the Americas, the question arose as to how you find gas leaks. In the desert the temperature ranges can be extreme, so the pipes undergo considerable expansion and contraction. It's uneconomic to send teams of people walking along the pipes looking for even pinhole leaks. But somebody had the bright idea of injecting a little ethanethiol into the gas stream. This is one of the sulphurous mercaptans that is enormously attractive to turkey vultures; they can detect even one or two parts per billion. Thus you get a column of the birds circling above even tiny leaks.
Imagine asking a computer to suggest some answer to the problem. It would be enough for it to reason that some agent needs to found which covers the territory, it doesn't have to be a single agent, and local birds would cover the territory adequately -- perhaps they can be involved. A large knowledge base wouldn't have to include any specialist information about the birds' powers of smell. It would be enough for it to suggest that there might be a way to lure the local birds to the scene of any leak; that would be enough of a clue to persuade you to go and find a suitable ornithologist to discuss the idea. A large knowledge base just has to include a huge amount of what is misleadingly called common knowledge. Beyond some critical mass it might well start to be useful for pointing humans in the right directions for many kinds of esoteric-looking problems.
Part of the moral here is that really clever systems are not those which know all the answers, they are systems which can ask good questions.
~What do you think AI tecnology will be like in 10 years? in 20? In 50? I hope there will be some new kinds of computing device, for a start. Von Neumann machines are limited; it would be rash to confine our ambitions to such architectures. Current neural net technology focusses unduly on simple units and feed-forward architectures, rather than units with (say) refractory periods and dynamic connectivity. So in the longer term I expect there will be more activity in what is loosely termed biocomputing.
There is an encouraging trend towards the study and application of complex dynamical systems, for example systems of simple autonomous devices that interact in ways that produce useful collective behaviour. At a guess, there will be some useful applications of such ideas -- even such mundane things as collections of small robots that together repair and maintain water and sewage distribution networks, a task which an ever-shrinking proportion of the world's communities can afford to do.
~Do you think Computers will ever be able
to think and talk like humans? Yes, but it's a long way off.
The triumphalism of some commentators is frightening; we've only really
just started to explore a few of the issues. But I'd say the same about
science in general -- although some people feel science has discovered most
of what's knowable, I'd say that we are still pretty much at the very beginning.
To praphrase an old joke: the ratio of what we know now to what we knew
a hundred years ago is truly enormous; but the difference is minute!
~What is the most exciting part of AI that
encourages
you stay in the field? Two things: the developing study of
complex dynamical systems, and the exploration of evolutionary computing
ideas. I doubt that any conventionally-designed system will ever have much
potential to be sufficiently adaptible to continue to function for any length
of time in the real world. Our design ideas are far too limited. The conventional
computing methodology of specify-design-implemnt may be fine for producing
virtually-bug-free software that satifies known needs, but it would be extremely
arrogant to suppose that we could ever set down the goals for any properly
intelligent system. We might be able to evolve such systems if we knew better
how to set about such a task. However, current GAs, even when used to develop
recurrent neural controllers or to develop robust computing algorithms,
are also pretty limited but there's a great deal of unexplored potential
in the field.
~What subjects would you encourage high-school
level students to take, who are interested in AI? Maths,
biology, some kinds of engineering, and perhaps a really good literature
course. The last of these can be an excellent way to learn to argue in a
truly versatile way and to learn how to express your ideas cleanly and convincingly.
Learn to communicate well with others!
~Other Comments: Unlike many
people, I'm rather skeptical about the internet as a medium for global knowledge-sharing.
Knowledge is power to many people, they regard it as just another saleable
commodity. Try using a good search engine such as Alta Vista to find out
about, say, a couple of topics that really matter to a lot of the world:
fleas, and cholera. Your search will turn up half a dozen pest-control firms
and nothing on cholera. There *is* some high-class technical information
available over the net on both these subjects, but you have to pay to get
it and you can hardly shop around for the cheapest deal because you don't
ever know what you've got until you've bought it. The current approaches
to commercialising the net show a depressing lack of imagination.
[ Basics | History | Approaches | Applications | Interactive | People | Links | Messages | Programs]