AI - The art and science of making computers do interesting things that are not in their nature.
This month, I've written another AI Alphabet. I hope - one of the entries will explain this - that after reading it, you will be thinking globally and breadth-first, not locally and depth-first.
Adi Shavit wrote to tell me about the LC++ library, developed by Brian McNamara and Yannis Smaragdakis, at www.cc.gatech.edu/~yannis/lc++/. The library uses macros to enable users to write in a Prolog-like syntax inside C++ programs. For example, here are some fragments of code from the tutorial:
FUN2(parent,string,string) FUN3(ancestor,string,string,int) DECLARE(Kid, string,2); DECLARE(Par, string,3); DECLARE(Anc, string,8); DECLARE(Tmp, string,9); DECLARE(X ,int, 10); DECLARE(Y ,int, 11); string bart="bart", lisa="lisa", maggie="maggie", marge="marge";, lassert( parent(marge,bart) ); lassert( parent(marge,lisa) ); lassert( parent(marge,maggie) ); lassert( ancestor(Par,Kid,1) -= parent(Par,Kid) ); lassert( ancestor(Anc,Kid,X) -= parent(Anc,Tmp) && ancestor(Tmp,Kid,Y) && X.is(plus,Y,1) ); iquery( ancestor(Anc,bart,X) );These declare the functors
ancestorand some logical variables and constants, assert a few rules, and then do a query.
The site includes a detailed paper about how LC++ is implemented.
"Proactive". "Synergetic". "Restructuring". "Leveraged"... And "Agent"? Does it mean anything, or is it just another buzzword? Franklin and Graesser apply linguistic philosophy to the word and propose a definition.
This short entry on the Conscious Entities site introduces a paper by Michael Anderson and Donald Perlis on brittleness, the inability of robots and programs to cope with unexpected developments. They mention a contestant in the DARPA Grand Challenge, where robots had to negotiate their way through a real-world journey. This one drove into a fence it could not see, and then continued trying to move forward for the rest of the contest. The robot was brittle. Sadly, the fence was not.
Suppose I am an AI agent in a video game: I have to shoot and disable a swarm of killer bees. Had I been designed around the principles of Good Old-Fashioned AI, the core of my cognition would be a propositional world-model continuously updated by interpreting sensory data, and I would represent each bee as a unique object within a world-centered coordinate system:
I see a bee. Sense data indicates I haven't seen it before. So I'll give it a new ID:
see( me, bee-12374 ). is_yellow( bee-12374 ). is_flying( bee-12374 ).
I turn my head. Then I turn back. I see a bee.
Is this new image I see also
bee-12374, or is it a different object to which I must allocate an ID?
And if the latter, how long must I retain my knowledge about
In contrast, deictic representations try to avoid problems of instance identification and world-model update by being relative to the agent and its intentions: this bee in front of me, which I am aiming at and intend to kill. The paper explains deictic representations, with references to important prior research, and proposes how the brain might use body movements such as changes in gaze position as deictic pointers.
Many people will know of Norman from his book The Design of Everyday Things. (The cover of one edition depicts a coffee pot with its spout on the same side as its handle.) In this Ubiquity interview on his The Future of Everyday Things, Norman explains why we should pay more attention to fun:
Norman: If you make something more pleasant, it's easier to use. (My first article on this subject is going to be published in the July/August 2002 issue of ACM's Interactions magazine.) The usability community has not paid enough attention to beauty, to fun, or to pleasure. I'd like to change that. The theme of affect and emotion is growing so much within me that I'm considering changing the title and focus of the book to "Emotion and Design."
Ubiquity: Incidentally, your own Web site, http://www.jnd.org/, is very nice, especially the trailing balls that follow cursor movement on the "Gratuitous Graphics" page. Do many Web sites achieve the right balance of fun and seriousness?
Norman: Not enough. For a lot of Web sites, it's all or none. It's rare to find a Web site that's really fun.
Emotion may be vital to AI programs as well as to their users. Without it, they may be unable to adapt correctly to threats and opportunities. As Norman explains,
the brain changes when we are happy, making pleasant objects easier to use. We are global, breadth-first thinkers when happy, local, depth-first thinkers when stressed. Affect is truly an important factor in how we live in the world. There are a lot of exciting new findings. I want to bring them to the attention of designers - and engineers who build large, complex, autonomous systems.
Giving a short explanation of the frame problem, this is one of many pages to quote the fable of R2D2 from Daniel Dennett's classic Cognitive wheels: The frame problem of AI:
Once upon a time there was a robot, named R1 by its creators. Its only task was to fend for itself. One day its designers arranged for it to learn that its spare battery, its precious energy supply, was locked in a room with a time bomb set to go off soon. R1 located the room, and the key to the door, and formulated a plan to rescue its battery. There was a wagon in the room, and the battery was on the wagon, and R1 hypothesized that a certain action which it called PULLOUT(WAGON, ROOM) would result in the battery being removed from the room. Straightaway it acted, and did succeed in getting the battery out of the room before the bomb went off. Unfortunately, however, the bomb was also on the wagon. R1 knew that the bomb was on the wagon in the room, but didn't realize that pulling the wagon would bring the bomb out along with the battery. Poor R1 had missed that obvious implication of its planned act.
Back to the drawing board. "The solution is obvious," said the designers. "Our next robot must be made to recognize not just the intended implications of its acts, but also the implications about their side-effects, by deducing these implications from the descriptions it uses in formulating its plans." They called their next model, the robot-deducer, R1D1. They placed R1D1 in much the same predicament that R1 had succumbed to, and as it too hit upon the idea of PULLOUT(WAGON, ROOM) it began, as designed, to consider the implications of such a course of action. It had just finished deducing that pulling the wagon out of the room would not change the colour of the room's walls, and was embarking on a proof of the further implication that pulling the wagon out would cause its wheels to turn more revolutions than there were wheels on the wagon . . . when the bomb exploded.
Back to the drawing board. "We must teach it the difference between relevant implications and irrelevant implications," said the designers, "and teach it to ignore the irrelevant ones." So they developed a method of tagging implications as either relevant or irrelevant to the project at hand, and installed the method in their next model, the robot-relevant-deducer, R2D1 for short. When they subjected R2D1to the test that had so unequivocally selected its ancestors for extinction, they were surprised to see it sitting, Hamlet-like, outside the room containing the ticking bomb, the native hue of its resolution sicklied o'er with the pale cast of thought, as Shakespeare (and more recently Fodor) has aptly put it. "Do something!" they yelled at it. "I am," it retorted. "I'm busily ignoring some thousands of implications I have determined to be irrelevant. Just as soon as I find an irrelevant implication, I put it on the list of those I must ignore, and . . ." the bomb went off.
"Natural selection is a mechanism for generating an exceedingly high degree of improbability". So runs Ronald Fisher's quote, heading this resource page for programs that work improbably well in finance - for example, Evolution of trading rules for the FX [foreign exchange] market, or, how to make money out of genetic programming. For those wanting to go back to the start of it all, the page links to Origin of Species and The Descent of Man.
Here's another AAAI page, a subtopic of their Resources for Students. Tips and suggestions deal with deciding on a topic and searching for articles, as well as typical questions: "could you please give me as much information on AI as possible"; "please send me any information that you have on Artificial Intelligence about the way it will affect the future"; "what are some threats and opportunities concerning artificial intelligence?"
In a robot arm, inverse kinematics is the problem of calculating what angles the joints need to be in order to get the end of the arm - the "hand" - to a desired position. This is a rather useful thing to know if the robot wants to reach towards and grab something. Forward kinematics - finding the hand position from the joint angles - is easy. Inverse kinematics is not, which is why a lot has been written about it, such as the nice article by Hugo Elias.
In Building Virtual Actors Who Can Really Act, Ken Perlin uses inverse kinematics to make game characters more plausible. Too many game characters move in exactly the same way whatever their psychological state. But as any animator knows, movement depends on mood; Wile E. Coyote may bound along expectantly as he's about to unwrap his new ACME Seed Your Own Tornados Kit, but he'll move rather differently when the tornados turn and come after him. Perlin's work involves fine-tuning or "shading" body movements to convey such subtleties. He does a lot of other things too, as the applets on his home page show. There's some amazing stuff there.
qsort( [one,two,three,four,five,six,seven,eight,nine], Sorted ).and back come the atoms sorted by name:
Sorted = [eight, five, four, nine, one, seven, six, three, two]
In the days when Sussman was a novice, Minsky once came to him as he sat hacking at the PDP-6. "What are you doing?" asked Minsky.
"I am training a randomly wired neural net to play Tic-Tac-Toe."
"Why is the net wired randomly?" asked Minsky.
"I do not want it to have any preconceptions of how to play."
Minsky shut his eyes.
"Why do you close your eyes?" Sussman asked his teacher.
"So the room will be empty."
At that moment, Sussman was enlightened.
John von Neumann said that "life is a process which may be abstracted from other media". If this is so, then just as we can study software independently of the hardware that runs it, we can study life independently of the matter that embodies it. One big question is how self-replicating molecules arose. Did it happen all at once, or over several stages? Is it logically necessary that if self-replicators arose elsewhere, they would do so in the "same" way, and if so, how can we characterise the kind of organisation involved? Artificial chemistry tries to answer such questions by setting up mathematical models of chemical reactions, keeping those properties needed to explain biological organisation while throwing away the rest.
One of the first artificial chemistry models, that of Fontana and Buss, was based on λ-calculus. Chemical reactions combine molecules to form new molecules, and we normally think of molecules as objects. But molecules are also functions: if given alcohol as an argument, water just mixes with it, but to sodium it does something more ferocious. (Certain chemists of my acquaintance used to dispose of sodium waste from their organic syntheses by tossing it into the Cherwell, presumably to frighten passing punts.) So we want a system whose elements are objects and at the same time functions that can act on these objects; which is what λ-calculus provides.
In 1999, MIT Provost Robert Brown asked a committee of faculty, students, and administrators to consider how the Internet would affect education, and how MIT should respond. The committee recommended that MIT should simply give its course materials away. With funding from the Andrew W. Mellon Foundation and the William and Flora Hewlett Foundation, the result was MIT's OpenCourseWare programme. To quote from one of many testimonials, from Maruf Muqtadir studying in Bangladesh:
Your OpenCourseWare is an amazing and remarkable step! I am currently a student of computer science at BRAC University of Dhaka, Bangladesh, and I find it very much useful to learn about my courses. I have always had a dream to study at MIT, since I came to know about the institution, its unique teaching methods, but for many reasons I am not able to do so. This initiative gives me the opportunity to self-teach myself.
Amongst the AI-related courses are Maths, Linguistics and Philosophy, Brain and Cognitive Sciences, and Electrical Engineering and Computer Science - including a downloadable textbook on inventions and patents, with a section on the future of American patents. There's much else too - Japanese language, sailing yacht design, US military budget and force planning - and more is to be added.
NetLogo, which my social-simulationist friend Edmund Chattoe recommended to me, is a free Logo designed for modelling systems containing very large numbers of similar agents. It has a lot of users and a lot of models built in it, and an active discussion list where new users can get advice.
I think that I shall never see
A matrix lovely as a tree.
Trees are fifty times as fun
As structures a la PL/I
Imagine you are an engineering student at an Oxford college (I mention no names) which admits students good either at engineering - the Gnomes - or at rugby - the Hearties. Suppose there's no correlation at all between being good at rugby and being good at engineering. However, as a Gnome, you are naturally biased against your beer-swilling rugby-playing fellows, and you decide to seek proof that their brains are abnormally small. You wander the grounds with your laptop and stats software, note the engineering and rugby ability of each co-student you pass, and then press "Compute Correlation". And Bingo!, you will see a negative correlation coefficient. (You can see why if you draw a scatterplot for all the students who applied to the college, chop out the area for those not admitted, and then consider the regression line for those left on the plot.)
This is Berkson's paradox, also known as "explaining away", and it's one of many topics discussed in Murphy's excellent paper on inference and learning with Bayesian networks, which also relates them to Kalman filters, Hidden Markov Models, and a number of other models. Bayesian networks have become popular in AI - as one researcher has said, they offer an efficient way to deal with the lack or ambiguity of information that has hampered previous systems, and provide an overarching graphical framework that brings together diverse elements of AI, increasing its range of application to the real world. Murphy, writing in 1998, says that the most widely used Bayes Nets are those embedded in Microsoft's products, including the Answer Wizard of Office 95, the Paperclip, and over 30 Technical Support Troubleshooters. Nowadays, they are being used in spam filtering.
I wonder whether Bullfighter, Deloitte's freeware program for detecting buzzwords, uses Bayesian filtering? Deloitte discovered a direct linkage between clear business talk and good business performance. In examining Enron's communications during its last three years, they found that as Enron began to sink, its press releases, financial reports, letters to shareholders, and speeches by top executives, became increasingly laden with ambiguous words and sentences.
"Qualia" is the philosophical term for subjective experiences: seeing red, feeling the sting of a wasp, tasting horseradish sauce. The Sony Qualia Man page claims that Sony take a close interest in qualia - as perhaps should any entertainment company - and that Ken Mogi, leader of their qualia project, has published a Qualia Manifesto calling for more of the things.
Many philosophers say qualia are the most important problem for the philosophy of mind. Why should I believe that you have subjective experiences? If you do, are they anything like mine, or could it be that whenever I see red, you see green? Will Artificial Intelligences have subjective experiences? If so, can we predict them from the nature of their programs? David Chalmers, who also wrote the Matrix as Metaphysics feature referenced in my Where am I? entry, discusses these problems in Absent Qualia, Fading Qualia, Dancing Qualia.
I predict that if we can ever know an AI's qualia, it won't be long before someone publishes a book on How to Make Your Computer See Red.
A recent entry in this excellent collection is Toddlers Sing With RUBI - two robots at UCSD are attending nursery school to teach songs, colors and shapes to one- and two-year old children. QRIO (for "Quest for Curiosity") from Sony, and RUBI (for "Robot Using Bayesian Inference"), developed at the Machine Perception Laboratory of UCSD, are there to study the uses of interactive computers for early childhood education.
Another entry recalls the rat-brained fighter pilot I featured last December. According to Lab Cultures Used to Create a Robotic 'Semi-Living Artist', researchers at the University of Western Australia and the Georgia Institute of Technology have created a new class of creative beings - a picture-drawing robot in Perth, Australia whose movements are controlled by the brain signals of cultured rat cells in Atlanta. They call it the semi-living artist.
Swarm Intelligence uses many simple agents to generate useful global behaviour via local interactions, no central controller needed. This site links to researchers, papers, software, and conferences. An interview with Eric Bonabeau, who has applied swarm intelligence to routing in telecoms systems, remembers how:
As a kid I'd always been terrified of insects. I remember with retrospective anguish my holidays in the south of France, when picnics turned into nightmarish fights against carnivorous wasps and ferocious ants raiding my sandwich. Sometimes I wonder how on earth I could dedicate eight years of my life to social insects. This large scale psychoanalytic phase transition took place in the early 1990's in Santa Fe, at the foot of the Rocky Mountains, the southernmost city before the New Mexican desert takes over. As a France Telecom R&D engineer, I was an unlikely candidate for such a radical transformation.
He continues by explaining that although one social insect may not be capable of much, a colony can achieve great things. A colony of ants can collectively find out where the nearest and richest food source is located, although no individual ant knows. If a food source is put near an ant nest, separated from it by a bridge with two branches, the colony is most likely to find the shorter branch. By laying and following pheremone trails, the ants perform an emergent computation, a route optimisation.
Participants in the original Turing test had to converse about topics such as arithmetic, weather and poetry; Shrager's computer must match a human in discussing photosynthesis. Amongst the links on this page are two excellent sites about the original Turing test.
Ubiquity is ACM's Web-based magazine, dedicated to fostering critical analysis and in-depth commentary, including book reviews, on issues relating to the nature, constitution, structure, science, engineering, cognition, technology, practices and paradigms of the IT profession. It published the interview with Donald Norman which I mentioned under E; and here is an interview with Stuart Russell, co-author with Peter Norvig of Artificial Intelligence: A Modern Approach. He has a crisp definition of AI:
An intelligent system is one whose expected utility is the highest that can be achieved by any system with the same computational limitations.
This is a nice paper on a course where final-year undergraduates at Bradford University were taught to build AIs for the real-time Artificial Life environment Terrarium Academic and the board game Virus. The authors aren't the first to teach via games - a famous example during the expert systems boom was Truckin', a game developed by Mark Stefik and others at Xerox Parc for teaching LOOPS. Indeed, I've done this too, with Traveller and Eden. So I'm not surprised to read that the Bradford students very much liked this style of Artificial Intelligence teaching, and that the authors hope to make freely available the clients and servers they built to enable different AIs to compete.
Chalmers wrote this paper for the philosophy section of the official Matrix website. As such, although most is intended for readers with no background in philosophy, it's a serious work, relevant to central issues in epistemology, metaphysics, and the philosophy of mind and language.
XSLT is the language developed for transforming XML documents into other XML documents. It's an interesting language, being entirely functional; and although we often think of XML as representing text, it can in fact represent general trees, so XSLT is a tree-transformation language. Ben-Kiki's posting links to an XSLT program for solving the N-queens problem (place N chess queens on an N × N square board so no queen threatens another); follow-ups suggest how XSLT could be improved.
There aren't too many AI-related words beginning with Y, so this letter called for a bit of searching. Simulated Sailing reviews sailing simulators. As Posey Yacht Design's pages explain - the review rated their Tactics and Strategy Simulator highly - even a stupid simulated opponent might have basic collision avoidance. But more intelligence is needed to handle matters such as getting clear air at the start, maintaining it against interference from other boats, and balancing against expected shifts in the wind.
Progressive Piss up at a yacht club comes from a slide presentation on constraint programming, using as an example a multi-boat party subject to complicated constraints on how crews circulate between boats. The objective is to minimise the number of boats. The slides stress how important it is to find the right formulation for constraint problems: increasing the search space may actually speed up search, if it reduces the number of variables and propagates constraints sooner.
With a contents which includes links to Agent-Based Computational Economics, the International Society of Artificial Life, and Alastair Channon's Evolutionary Emergence of Intelligent Behaviours, not to mention Algorithmic Chemistry, this is an excellent starting point for Artificial Life explorations. Though if you believe the statistical argument in Are You Living in a Computer Simulation?, we're probably already Artificial Life.
Getting back to the history lesson, the prospects for the decade look mostly medical. Progress is expected to speed up shortly, as the fundamental patents in genomic engineering begin to expire: the Free Chromosome Foundation has already published a manifesto calling for the creation of an intellectual-property free genome with improved replacements for all commonly defective exons.
Experiments in digitizing and running neural wetware under emulation are well-established; some radical libertarians claim that as the technology matures, death - with its draconian curtailment of property and voting rights - will become the biggest civil rights issue of all.
Some commodities are expensive: the price of crude oil has broken sixty euros a barrel and is edging inexorably up. Other commodities are cheap: computers, for example - hobbyists print off weird new processor architectures on their home inkjets; middle aged folks wipe their backsides with diagnostic paper that can tell how their VHDL levels are tending.
The latest casualties of the march of technological progress are: the high street clothes shop, the flushing water closet, the Main Battle Tank, and the first-generation of quantum computers. New with the decade are cheap enhanced immune systems, brain implants that hook right into the Chomsky organ and talk to you using your own inner voice, and widespread public paranoia about limbic spam. Nanotechnology has shattered into a dozen disjoint disciplines, and skeptics are predicting that it will all peter out before long. Philosophers have ceded qualia to engineers, and the current difficult problem in AI is getting software to experience embarrassment.
Fusion power is still, of course, fifty years away.
Quoted from Tourist by Charles Stross: originally published in Isaac Asimov's Science Fiction Magazine for February 2002, republished 2005 in his novel Accelerando, and downloadable under the Creative Commons License from www.accelerando.org/.[ Jocelyn Ireson-Paine's Home Page ]