Sections:
|
Tuesday, October 14, 2008
How Spam is Improving AI
Anti-spam puzzles are helping researchers develop smarter algorithms.
By Kurt Kleiner
Those pesky visual puzzles that have to be completed each time you sign up for a Web mail account or post a comment to a blog are under attack. It's not just from spam-spewing computers or hackers, though; it's also from researchers who are using anti-spam puzzles to develop smarter, more humanlike algorithms. The most common type of puzzle (a series of distorted letters and numbers) is increasingly being cracked by smarter AI software. And a computer scientist has now developed an algorithm that can defeat even the latest photograph-based tests. Known as CAPTCHAs (Completely Automated Public Turing test to tell Computers and Humans Apart), these puzzles were developed in the late '90s as a way to separate real users from machines that create e-mail accounts to send out spam or log in to message boards to post ad links. The Turing Test, named after mathematician Alan Turing, involves measuring intelligence by having a computer try to impersonate a real person. Textual CAPTCHAs are a good way to tell humans and spam-bots apart, because distorted letters and numbers can easily be read by real people (most of the time) but are fiendishly difficult for computers to decipher. However, computer scientists have long seen CAPTCHAs as an interesting AI challenge. Designers of textual CAPTCHAs have gradually introduced more distortion to prevent machines from solving them. But they have to balance security against usability: as distortion increases, even real human beings begin to find CAPTCHAs difficult to decipher. Earlier this year, Jeff Yan, a researcher at the University of Newcastle, U.K., revealed a program capable of completing the textual CAPTCHAs used to protect Microsoft's Hotmail, MSN, and Windows Live services with a success rate of 60 percent. This might not sound like much, but it's significant, since a computer can try its attack thousands of times each minute. Yan withheld the paper until Microsoft had a chance to tweak its CAPTCHAs so that they were more difficult to crack. But at the ACM Computer and Communication Security Conference in Alexandria, VA, later this month, Yan will present details of another program that he says can crack even more widely used textual CAPTCHAs. So an alternative is to ask users to solve different kinds of puzzles. But another paper to be presented at the same conference describes an algorithm that could spell trouble for even newer CAPTCHAs. Philippe Golle of the Palo Alto Research Center has developed a program that can correctly pass an image-based CAPTCHA called Asirra, developed by Microsoft. Asirra asks users to correctly classify images of either cats or dogs using a database of three million images provided by animal-rescue organizations. This task should be even harder for computers than recognizing squiggly letters, but Golle's program can correctly identify the cats or dogs shown by Asirra 83 percent of the time. Golle trained his program using 8,000 images collected from the same website. Through trial and error, his software gradually learned to tell cats and dogs apart, based on a statistical analysis of color and texture in each photo. The pink of the dogs' tongues and the green of the cats' eyes provided strong clues, Golle says, but it is only by studying color and texture information from so many images that his program could attack the problem. "Machine learning is very good at aggregating information," Golle says. However, although each individual picture was recognized 83 percent of the time, the full CAPTCHA test requires 12 pictures to be identified simultaneously, so the attack actually works only 10.3 percent of the time. Golle says that an easy countermeasure would be for Asirra to present more pictures, which would further drive down the success rate of the attack. Microsoft did not respond to our requests for comment. Despite all this progress, it's unclear whether or not real spammers are currently using AI attacks against real CAPTCHAs. Websense Security Labs, in San Diego, has released reports about spammers cracking CAPTCHAs, but often this involves simply having low-paid workers solve CAPTCHAs manually. Luis von Ahn, a computer scientist at Carnegie Mellon University, who helped coin the term CAPTCHA, says that it's not clear that any common CAPTCHAs have been broken by machine attack in the real world. "I don't know of anybody who's thinking of getting rid of the CAPTCHA because it doesn't work," he says. However, von Ahn notes that using humans comes at a cost. Even if workers are paid just $3 per 1,000 CAPTCHAs, that is expensive, he says, especially since most of the hacked Web mail accounts will be shut down soon after they begin to send out spam. So a truly automated attack would reduce the cost to spammers and greatly increase the number of successful attacks they could afford, he says. But until computers start to get much smarter, CAPTCHA creators will always be able to implement a few simple tweaks to make a CAPTCHA much harder. "I do think there will be a day when, essentially, CAPTCHAs are going to be useless," von Ahn says. "But I don't think it's this year, or next." Copyright Technology Review 2008. Op-Ed Contributor - The Rise of the Machines - NYTimes.com
www.nytimes.com/2008/10/12/opinion/12dooling.html?... October 12, 2008
Op-Ed Contributor
The Rise of the MachinesBy RICHARD DOOLING
Omaha “BEWARE of geeks bearing formulas.” So saith Warren Buffett, the Wizard of Omaha. Words to bear in mind as we bail out banks and buy up mortgages and tweak interest rates and nothing, nothing seems to make any difference on Wall Street or Main Street. Years ago, Mr. Buffett called derivatives “weapons of financial mass destruction” — an apt metaphor considering that the Manhattan Project’s math and physics geeks bearing formulas brought us the original weapon of mass destruction, at Trinity in New Mexico on July 16, 1945. In a 1981 documentary called “The Day After Trinity,” Freeman Dyson, a reigning gray eminence of math and theoretical physics, as well as an ardent proponent of nuclear disarmament, described the seductive power that brought us the ability to create atomic energy out of nothing. “I have felt it myself,” he warned. “The glitter of nuclear weapons. It is irresistible if you come to them as a scientist. To feel it’s there in your hands, to release this energy that fuels the stars, to let it do your bidding. To perform these miracles, to lift a million tons of rock into the sky. It is something that gives people an illusion of illimitable power, and it is, in some ways, responsible for all our troubles — this, what you might call technical arrogance, that overcomes people when they see what they can do with their minds.” The Wall Street geeks, the quantitative analysts (“quants”) and masters of “algo trading” probably felt the same irresistible lure of “illimitable power” when they discovered “evolutionary algorithms” that allowed them to create vast empires of wealth by deriving the dependence structures of portfolio credit derivatives. What does that mean? You’ll never know. Over and over again, financial experts and wonkish talking heads endeavor to explain these mysterious, “toxic” financial instruments to us lay folk. Over and over, they ignobly fail, because we all know that no one understands credit default obligations and derivatives, except perhaps Mr. Buffett and the computers who created them. Somehow the genius quants — the best and brightest geeks Wall Street firms could buy — fed $1 trillion in subprime mortgage debt into their supercomputers, added some derivatives, massaged the arrangements with computer algorithms and — poof! — created $62 trillion in imaginary wealth. It’s not much of a stretch to imagine that all of that imaginary wealth is locked up somewhere inside the computers, and that we humans, led by the silverback males of the financial world, Ben Bernanke and Henry Paulson, are frantically beseeching the monolith for answers. Or maybe we are lost in space, with Dave the astronaut pleading, “Open the bank vault doors, Hal.” As the current financial crisis spreads (like a computer virus) on the earth’s nervous system (the Internet), it’s worth asking if we have somehow managed to colossally outsmart ourselves using computers. After all, the Wall Street titans loved swaps and derivatives because they were totally unregulated by humans. That left nobody but the machines in charge. How fitting then, that almost 30 years after Freeman Dyson described the almost unspeakable urges of the nuclear geeks creating illimitable energy out of equations, his son, George Dyson, has written an essay (published at Edge.org) warning about a different strain of technical arrogance that has brought the entire planet to the brink of financial destruction. George Dyson is an historian of technology and the author of “Darwin Among the Machines,” a book that warned us a decade ago that it was only a matter of time before technology out-evolves us and takes over. His new essay — “Economic Dis-Equilibrium: Can You Have Your House and Spend It Too?” — begins with a history of “stock,” originally a stick of hazel, willow or alder wood, inscribed with notches indicating monetary amounts and dates. When funds were transferred, the stick was split into identical halves — with one side going to the depositor and the other to the party safeguarding the money — and represented proof positive that gold had been deposited somewhere to back it up. That was good enough for 600 years, until we decided that we needed more speed and efficiency. Making money, it seems, is all about the velocity of moving it around, so that it can exist in Hong Kong one moment and Wall Street a split second later. “The unlimited replication of information is generally a public good,” George Dyson writes. “The problem starts, as the current crisis demonstrates, when unregulated replication is applied to money itself. Highly complex computer-generated financial instruments (known as derivatives) are being produced, not from natural factors of production or other goods, but purely from other financial instruments.” It was easy enough for us humans to understand a stick or a dollar bill when it was backed by something tangible somewhere, but only computers can understand and derive a correlation structure from observed collateralized debt obligation tranche spreads. Which leads us to the next question: Just how much of the world’s financial stability now lies in the “hands” of computerized trading algorithms? • Here’s a frightening party trick that I learned from the futurist Ray Kurzweil. Read this excerpt and then I’ll tell you who wrote it: But we are suggesting neither that the human race would voluntarily turn power over to the machines nor that the machines would willfully seize power. What we do suggest is that the human race might easily permit itself to drift into a position of such dependence on the machines that it would have no practical choice but to accept all of the machines’ decisions. ... Eventually a stage may be reached at which the decisions necessary to keep the system running will be so complex that human beings will be incapable of making them intelligently. At that stage the machines will be in effective control. People won’t be able to just turn the machines off, because they will be so dependent on them that turning them off would amount to suicide. Brace yourself. It comes from the Unabomber’s manifesto. Yes, Theodore Kaczinski was a homicidal psychopath and a paranoid kook, but he was also a bloodhound when it came to scenting all of the horrors technology holds in store for us. Hence his mission to kill technologists before machines commenced what he believed would be their inevitable reign of terror. • We are living, we have long been told, in the Information Age. Yet now we are faced with the sickening suspicion that technology has run ahead of us. Man is a fire-stealing animal, and we can’t help building machines and machine intelligences, even if, from time to time, we use them not only to outsmart ourselves but to bring us right up to the doorstep of Doom. We are still fearful, superstitious and all-too-human creatures. At times, we forget the magnitude of the havoc we can wreak by off-loading our minds onto super-intelligent machines, that is, until they run away from us, like mad sorcerers’ apprentices, and drag us up to the precipice for a look down into the abyss. As the financial experts all over the world use machines to unwind Gordian knots of financial arrangements so complex that only machines can make — “derive” — and trade them, we have to wonder: Are we living in a bad sci-fi movie? Is the Matrix made of credit default swaps? When Treasury Secretary Paulson (looking very much like a frightened primate) came to Congress seeking an emergency loan, Senator Jon Tester of Montana, a Democrat still living on his family homestead, asked him: “I’m a dirt farmer. Why do we have one week to determine that $700 billion has to be appropriated or this country’s financial system goes down the pipes?” “Well, sir,” Mr. Paulson could well have responded, “the computers have demanded it.” Richard Dooling is the author of “Rapture for the Geeks: When A.I. Outsmarts I.Q.” daedalus-winter2004.pdf (application/pdf Object)
www.chimera.info/daedalus/downloads/daedalus-winte... Daedalus Winter 2004 - Special Issue on Role-Playing Games
MAKE: Blog: iRobot - Photos and video of the PackBot!
www.makezine.com/blog/archive/2006/06/irobot_photo...
FSU researcher's work on unmanned ground vehicles could save soldiers' lives
www.fsu.edu/news/2006/06/01/unmanned.vehicle/
Facial Recognition Software VeriLook can be easily integrated into the customer's security system. The integrator has a complete control over SDK data input and output; therefore SDK functions can be used in connection with most cameras (including webcams) and any database. Integrator could develop any user interface. VeriLook is available as VeriLook 3.0 Standard SDK. VeriLook 3.0 Standard SDKVeriLook 3.0 Standard SDK allows to develop face identification systems for Microsoft Windows, Linux and Mac OS X platforms. The kit includes programming samples in several programming languages.  Click to zoom  Click to zoom  ArsGeek » Blog Archive » Give Me Some Rubber, Glue and I Can Take Over Your Life!
arsgeek.com/?p=2258 Biometrics - your face is your password
In a radical new approach to solving identiy theft, CBL researchers are using three-dimensional information to obtain a unique biometric signature of a person's face. With cutting-edge hardware and novel algorithms, they are designing system that turns a process practically as effortless as taking a photograph into a powerful authentication protocol. Remembering dozens of personal identification numbers and passwords is not the solution to identity theft. Both are inconvenient to memorize and impractical to safeguard, and in essence merely tie two pieces of information together. Once the secret is compromised, the rest follows. The solution is to be able to tie private information to its owner in a way that cannot be compromised - biometric authentication The CBL's URxD system has the potential to move face recognition technology to the high performance gear needed for widespread application. The system determines not only the characteristics of each face, but also whether the person is wearing glasses, allowing for a practical system which offers high accuracy. So far, face recognition methods have focused on appearance - capturing, representing, and matching facial characteristics as they appear on two-dimensional images in the visible spectrum. This is quite challenging to machine recognition because such characteristics vary with orientation, age, habits (e.g., bearded appearance), and illumination. Instead, our system uses three-dimensional information, and has achieved the best published results when tried to 4,007 datasets (part of the international face recognition Grand Challenge organized by NIST). These results show strong promise in overcoming the difficult problems that have been holding back progress in this field for many years. Official Google Blog: A better way to organize photos?
googleblog.blogspot.com/2006/08/better-way-to-orga...
AIIDE (AI and Interactive Digital Entertainment) AIIDE is the definitive point of interaction between entertainment software developers interested in AI and academic and industrial AI researchers. Sponsored by the American Association for Artificial Intelligence (AAAI), the conference is targeted at both the research and commercial communities, promoting AI research and practice in the context of interactive digital entertainment systems with an emphasis on commercial computer and video games. AIIDE-07 author registration and submission index site http://www.aaai.org/Conferences/AIIDE/aiide07.php. IMPORTANT DATES December to January: Authors register on the AIIDE web site January 22, 2007: Electronic submission of full paper February 2, 2007: Electronic submission of demonstration abstract March 10, 2007: Notification of acceptance decision iRobot Scooba Exposed: What's Inside This Robotic Maid > Disassembly
www.informit.com/articles/article.asp?p=474235&rl=...
Bootstrapping the brain: unsupervised program learns baby talk
arstechnica.com/news.ars/post/20070724-bootstrappi... Bootstrapping the brain: unsupervised program learns baby talkBy John Timmer | Published: July 24, 2007 - 11:31PM CT Nature-versus-nurture debates have been going on for centuries, but they have been steadily growing in sophistication as we learn more about how the brain functions and appreciate how self-organizing systems can produce elaborate behavior. One of the areas where the debate plays out is in the acquisition of language, where cognitive scientists have been arguing over whether the brain comes with some preset patterns that it fits language into, or whether it generates its entire language recognition system on the fly. In an attempt to help address this issue, some researchers have explored the ability of systems to generate their own language categorization system and, in the process, crafted a program that can learn to recognize vowels on its own. The work built on previous studies that had suggested that children are best able to categorize speech when different tokens are presented to them as a distinct peaks along a Gaussian curve. So they devised algorithms that could identify higher-dimension Gaussian clustering of speech tokens. These systems then trained themselves with vowel sounds derived from recordings of people speaking to babies in either Japanese and English. After a few thousand iterations, they were tested for accuracy against further recordings. The key feature of these tests was that the programs were not given any information as to how many different vowel sounds would be presented (it was four in these tests). The first system they tried took what might be viewed as an evolutionary approach. It was simply supplied with a large number of randomly generated Gaussian clusters and scored each cluster based on how well they encompassed the data that they were fed. Over many iterations, clusters that provided poor fits were eliminated, ultimately leaving just a few that fit the data best. The results were impressive: the majority of the models correctly recognized four distinct tokens, and could identify them with over 90 percent accuracy. Unfortunately, as the authors noted, the brain is unlikely to come with a set of predefined Gaussian clusters to select from, so this program was probably a poor representation of the actual process of learning language. So, they set up a second algorithm, this one closer in structure to a neural network in a three-dimensional space. If a given speech token fell at a given node, the connections between it and the surrounding nodes were strengthened. Here, the results weren't as good, but the models did recognize the four vowel tokens with over 80 percent success rates. The work has some implications for the nature-versus-nurture debate here, as it suggests that the brain doesn't necessarily need preformed language capacity but may require the ability to recognize Gaussian distributions. This puts the results somewhere between the two extremes. The authors note that other results had already suggested the brain had this capacity, both for language and other recognition processes. But the results also have some interesting implications for self-training computer systems. Given a few basic audio properties, the evolutionary system could train itself to both identify the number of categories they represent, and successfully place other speech tokens into the correct categories. It's possible that similar approaches could work for other recognition processes, such as visual identification.
PNAS has released the paper to the press, but has yet to make it available to subscribers. When it is released, it should appear here.
Engineering Spore
By David Kushner
It's time to make a creature. Let's start with its body. Stretch down a pair of legs and pull out two arms so that it looks long and lean. On one end of the body, pop on intelligent eyes behind large round glasses. Add a mop of peppery hair and a prominent nose and ears. Sprinkle on scruffy semibeard growth. Call it Will Wright. Now put the creature in its habitat: the workspace of a computer game developer in Emeryville, Calif. Spread the studios of Maxis Software, which Wright cofounded in 1987, over the floors of two nondescript office buildings. Sprinkle the interior with dirt-encrusted mountain bikes leaning against cubicle walls and overgrown, pumpkin-orange beanbag chairs. Now surround Wright with others of his kind, hunched behind desks, typing at keyboards, PC monitors glowing. The software engineers, artists, and others who work at Maxis, now owned by video-game giant Electronic Arts, have migrated here because Wright is a legend. Over the past two decades, the 48-year-old Wright, who studied architecture and mechanical engineering at Louisiana Tech University, has utterly transformed his industry with hits like SimCity, in which players build virtual towns, and the best-selling computer game franchise of all time, The Sims, in which players create virtual people and then watch them interact. In the process, Wright has helped forge a new, more toylike frontier in computer gaming, where the main goal is not so much to score points or kill bad guys but to create cool stuff. The game they're working on this bright February day is called Spore, and it's the most ridiculously ambitious simulation game yet. Sure, there've been virtual worlds like Second Life, which let you customize your characters. In games like Fable and Black & White, the characters even evolve in appearance and reputation based on how the player defines them: the more evil your beast, for example, the more feared it becomes. LittleBigPlanet, an upcoming game from Media Molecule for Sony's PlayStation 3 game console, is built around player-made terrains and characters. But Wright's Spore is by far the boldest in terms of unleashing players' creativity. In Spore the players create life itself—starting with ooze-dwelling, one-celled creatures that learn, grow, and evolve into intelligent beings with advanced cultures and technologies, able to conquer their planets and outer space. Computer gamers everywhere have eagerly awaited Wright's latest project since he began talking about it in 2000. Spore is finally due to be released this month, more than a year behind schedule. Wright attributes its recent delays to localization, the process of tailoring the game to different countries and languages. Others around the Maxis office cite the boss's high expectations. Wright concedes their point but shrugs it off. “For games, it is a long time, but for me it's not a big deal,” he says, sipping coffee in his cluttered corner office. “I'd rather spend a couple of extra years and have it be a big seller than short it by a year or two and have it be mediocre.” Spore is anything but. Other games may look and sound better, but few games are as original as this one. It offers players far more choice and open-ended play than any game before it. If Spore lives up to its creator's vision, it will likely be heralded as one of those milestones that redefines what a game can be—just as Doom, a first-person shooter game, pioneered fast-action multiplayer competition in 1993 and Guitar Hero delivered the thrill of music performance by introducing a guitar-shaped controller.
The anticipation—and pressure—is high. “I call Wright a genius because he truly is one of the most innovative developers out there,” writes one gaming blogger. “Spore...is creating an entirely new genre.” The game unfolds through five stages, each of which riffs on an established genre of play. It starts, fittingly, in a two-dimensional world, with a single-celled organism that gobbles up microbes and plants to accrue DNA points. Once the spore collects enough DNA, an editing palette pops up that lets you design the next evolutionary stage of the creature's body. Your creature is then thrown into a three-dimensional environment where it must dodge predators and find a partner with which to reproduce. By the third stage, your creature is fully evolved and you switch to controlling its entire tribe, as you would in real-time strategy games like Electronic Arts' Command & Conquer. Next up is the civilization phase, in which you can assemble vehicles and buildings to bring your tribe's city to life, in the spirit of SimCity or Civilization. If you succeed in conquering your planet and avoiding an enemy takeover, you graduate to the fifth and final level: outer space. Here the object is to fight off invaders and take over other planets. Developing a game in which the players create all the key parts—the characters, buildings, and vehicles—poses an obvious conundrum: “There's no content,” says Maxis technology fellow Chris Hecker. “Initially, the problem was, well, what is [Spore] supposed to be?” When you boot up most games for the first time, you're immediately immersed in an existing world, complete with a cast of characters who behave in predetermined ways. Perhaps the game has tree-lined streets or castles with dungeons and moats. Maybe colonies of dwarves and trolls populate those worlds, or maybe gangsters do. These objects are all encoded in the game's original software exactly as the developers envisioned and animated them. In Spore, that model doesn't apply. Almost nothing exists until the player makes choices about each object's shape and texture. To enable that design process, the relatively small team of 20 artists and seven programmers created a palette of editing tools. They think of it as an “artist in a box” or, as Maxis software engineer Colin Andrews puts it, “Mr. Potato Head on steroids.” Maxis released a stripped-down version of the tool palette, known as the Creature Creator, on 16 June to build buzz for the game. Eight days later, early adopters had created more than a million creatures. To understand why the Creature Creator is so compelling, consider its incredible flexibility. Say a player wants to make a building. Spore provides a menu of architectural elements to tinker with: windows, doors, that kind of thing. The player clicks and drags the pieces onto a base structure and can stretch or shrink them along several axes. From the game's perspective, each building's design is simply a list of instructions; when the player is finished tinkering, those instructions direct the game engine to generate an image of the building and place it within the Spore world. Simple enough. Then there's the process of making a creature, which offers a whole other level of variety and complexity. For instance, each creature can have any number of features and appendages—eyes, mouths, legs, feet—which can be stretched and curled like clay into outlandish shapes. But that indeterminacy presents an unusual problem: how exactly does a game company write software that generates realistic movements for “an eight-legged, two-headed thing with four mouths and no neck?” Wright says. “We don't know what we're animating.” To convincingly evoke even the wackiest animal a player could design, the game code had to be able to apply the knowledge of a human animator, on the fly—the ability to understand body language and subtle facial expressions and then to encapsulate those qualities abstractly in software. Wright decided to build Spore's real-time animation around a technique called procedural generation. The “procedure” in Spore is a set of algorithms that execute a player's designs, generating entirely new content in the midst of game play. Other game developers have used the technique for years in a limited way, but no game has ever relied on it so heavily to create highly customizable yet lifelike creatures in real time. So the Maxis team had little to go on as they tried to figure out how to make their exotic beasts move. Wright, who builds BattleBots for fun and possesses a voracious intellect and curiosity, decided to hit the books. He began reading up on biomechanics—in particular, the physics and physiology of how animals move. “Depending on the leg length and how supple the spine is,” he says, “you can get a characteristic oscillation of the [torso] of the creature over the ground.”
To get a creature to walk or run convincingly, the software engineers encoded an overarching set of rules on how to generate movement. The animation algorithms start by looking at the number of legs, the length of each leg, and the creature's bodily symmetry to calculate something called a walking ratio. If one side has twice as many legs as the other, for example, the ratio would be one to two. The algorithms will also compute the rhythm of a creature's footfalls—the length of time between, say, a front leg and a back leg hitting the ground in a single stride. The overall gait takes all these factors into account, along with the dimensions of the torso and head (or heads, as the case may be). The result is a convincingly lifelike motion. The Maxis team then had to see if those movement rules worked on actual Spore beings. The team devised a huge menagerie of test creatures, observed how they stumbled around, and then adjusted the software's algorithms, essentially creating a virtual island of Dr. Moreau. “We have these tests where we take 10 totally crazy, random creatures and run animation on them,” Wright says. “And we find out that it works for these seven, but for these two the legs look weird and for this one the back isn't straight enough. We're refining those algorithms all the time.” To illustrate, Wright goes over to his Dell computer and, with a few pecks at the keyboard, brings up the game's Creature Creator on screen. He starts with a short, fat torso and attaches birdlike wings on its sides for ears. From a palette of eyeballs, he clicks and drags a pair of big round eyes and drops them onto the beast's shoulders. He continues to tweak the anatomy, equipping the creature with legs, arms, hands, and so on. Even as Wright experiments with different looks, the beast begins to move—wiggling its newly attached limbs and blinking its new eyes. It even seems to show its approval of certain choices by smiling and nodding. Based on each new creature's features and shape, the animation software determines the sounds it can make, the way it dances, and much more. A skinny beast with a beak and decorative tufts of hair may flutter its eyelashes and emit a high-pitched warble, while a hulking creature with spikes along its spine may blink slowly and communicate in a baritone growl. Those traits in turn end up influencing whether the creature greets other species with a friendly advance or with an attack, and the fate of its civilization depends on those nuances. Wright clicks on a button to test the creature he's just designed. With short legs on one side and long legs on the other, the animal lumbers awkwardly but convincingly across the screen. Indeed, the little legs scurry just fast enough to keep up with the long ones. But Wright isn't done with him yet. “What would it take to make any creature sad?” he asks, tapping away at the keyboard. Suddenly this alien being adopts a recognizably sad pose, dropping its torso, curling down the edges of its mouth, and dully drooping its eyelids. You feel kind of sorry for the little guy. This beast is relatively straightforward, but the Maxis team had to allow for the most twisted possibilities a player might dream up—for instance, a creature with no limbs. “Now the game has to deal with all the ramifications of that,” says Hecker. “So how do you pick up a piece of fruit?” In conventional animation systems, the concept of a limb may be encoded not as an object but rather as a set of spatial transformations that can be applied to a body. To accomplish this, an animator can assign labels to parts of a character's skeleton. When a character reaches for that fruit, the animation might state something like, “Rotate bone 1 from 0 to 52 degrees.” But in Spore the skeleton is unknown until the game is already in play. So instead of using labels, the programmers encode generic descriptions of each body part, referring to a specific limb by describing its context relative to other body parts. Let's say a creature throws a punch at a bad guy. The animation may dictate the action with instructions that would read something like, “Move upper leftmost grasper from rest position to a position parallel to your leftmost head, then move to some position relative to the enemy's topmost nose.” The code analyzes the body in search of the parts that match that profile. In short, rather than directing bones to assume prescribed positions, animators are using higher-level directives to describe what the bones should do. This strategy was key to opening up a much wider field of character types and activities—though it certainly didn't make writing the game easy. The code can look for a limb by using a description that may be satisfied by one body part, several, or none. Coping with indeterminate results, while keeping the animation interesting by not simply ignoring extra limbs, drives up the complexity of the game code. And then there are the creatures that are so weird they defy the game's generalized rules of movement. To catch those freakish cases, the code checks for certain features: for example, does the creature have any kind of paw or claw? If not, a separate set of instructions will govern the creature's movement, instructing it to use its mouth as a hand. And because the creature has no legs to use in calculating movement, it gets an inchwormlike slither. Rather than try to write very complex algorithms that cover every imaginable kind of beast, the programmers instead identified a few exceptional creature skeletons and wrote code that chooses different sets of rules for them. “It just kind of ripples out in a lot of different ways,” Hecker says. It's one thing to create and animate a creature in Spore, but Wright and his team knew that players would also want to share their creations. “What we saw with The Sims was that people loved downloading tools and creating stuff in the game,” says Wright. Players routinely surfed Web sites specially set up for trading Sims add-ons, such as modifications to a character's appearance, houses, and furniture. But the experience was a hassle because players had to find and then import each item into the game one by one. “We wanted to basically make that [process] part of the game play,” he says. To do that, Maxis devised the Pollinator. This tool lets players easily search through the buildings, vehicles, and beasts created by other gamers and incorporate them into their own worlds. They can also sort through the stuff by theme—whimsical or Wizard of Oz, futuristic or “Futurama.” This is what Wright means when he describes the game as “a massively single-player experience”: it's a one-person game that can draw on many sources. While the Spore DVD will ship with some ready-made creatures and buildings—so that a player's creatures aren't initially running around in an empty universe—the rest of the content will come from the players, who can upload their creations to Maxis's game servers for others to access. To store and sift through such a huge amount of data, the Maxis team had to compress the data files down to a manageable size. Here, the hurdle became “how do we keep the data rate really low so that even if I'm not on the Internet I can still have the local database with lots and lots of content?” Wright says. With its detailed terrains and texture maps, one planet in Spore could occupy 10 megabytes of space on a player's hard drive. “We don't have the disc space to deliver a million planets,” says Maxis art director Ocean Quigley. When a player creates a planet, an instruction list for generating that world is saved along with certain seed values, which are like keys that the software uses to reopen the world later. To conjure up lots of different planets for each game, the code requires that certain values or quantities be more or less randomly assigned. Algorithms embedded in the game's software can generate those strings of seemingly unrelated numbers, but the starting value—or seed—must vary so as to avoid generating the same string of numbers each time the algorithms are run. “We want the planets to be ‘random,' but we also need to be able to re-create it exactly when you come back later,” Andrews explains. “Storing the seed lets us do that.” The programmers also had to winnow down the list to just the core guidelines needed to reconstitute a planet—or building, creature, or spaceship. Sometimes that meant making tough choices that in effect curtail a player's creativity. Originally, for instance, Spore's Creature Creator allowed players to design animals with looped spines. Unfortunately, doughnut-shaped animals raised all sorts of exceptions to the animation rules. The solution: bye-bye, doughnuts. To manage the flow of so many player-created creatures and items, and to help players find content they like, Spore uses the same kind of collaborative filtering that sites like Amazon and Netflix have made popular, based on the preferences of other players who have chosen a certain design. Players will also be able to subscribe to Sporecasts—a kind of RSS feed of content other people create for the game. As Spore spreads, stars of content design will likely emerge, as they have in Second Life and in other online gaming communities. “I can imagine so many cool possibilities that we're just scratching the surface of,” Wright says. He envisions Spore races centered around user-designed vehicles and flying games featuring users' spaceships. But Wright's imagination stretches only so far. He anticipates the day when Spore players take charge and steer the game into unseen territory. “That's when the fans become an even larger designer [than us],” he says. “In some sense, we're kind of codesigning Spore. Fans are going to drive its future.” Computer program to take on the Unabomber | The Register
www.theregister.com/2007/07/24/man_machine_poker_c... Computer program to take on the UnabomberMan vs. Machine challengePublished Tuesday 24th July 2007 05:11 GMT Mobile computing: Opportunities and risk - Free whitepaper
Wireless Email Solutions - Free whitepaper');}
How intelligent is a computer - and just what does it mean to be intelligent, anyway? The Association for the Advancement of Artificial Intelligence (AAAI) aims to clarify that question with its Man vs. Machine Poker Challenge, taking place today and tomorrow in Vancouver, British Columbia. Polaris, a poker program developed at the Computer Poker Research Group of the University of Alberta, will compete against poker professionals Phil “The Unabomber” Laak and Ali Eslami in a two day Texas Hold’em tournament designed to test the limits of computer “thought” in real world scenarios. Part of the challenge for proponents of artificial intelligence lies in the difficulty of defining exactly what intelligence is: “intelligence” as we typically understand it is a rather haphazard measurement of our abilities, some of which lend themselves to duplication via computer software, and some of which do not. Human intelligence is as much about communication and nuance as it is about performing mathematical functions, which is all that computers are really good at. Computers, in short, excel at the purely computational tasks that lend themselves to ready measurement, which is why within the closed universe of the chessboard computers are capable of beating even the greatest human players. On the other hand, computers do miserably at the forms of emotional intelligence that define so much of what we are as human beings. Poker - a game of imperfect information, with its potent mixture of probability and human psychology – is the kind of talent that in the past has humbled even the most powerful supercomputers. Polaris seeks to redress that imbalance. Two copies of the program will compete as a team against the two professionals in a version of duplicate poker, a poker variant designed to minimize the element of chance. In duplicate poker, players at different tables play the same hand and compete not against those at their own table, but against those playing the same hand at different tables. Luck is still involved, inasmuch as no one can predict the behavior of the other table’s opponents, but the game is largely one of skill, which is why it has slipped under the radar of the ever-vigilant American authorities.
In the Man vs. Machine Poker Challenge, each teammate will play the hand of his teammate’s opponent, and teammates are forbidden to communicate with each other during the match. The match should therefore provide a fairly accurate gauge of the relative skill of each individual player, and provide at least a glimpse of just how much of that emotional intelligence can be coded into computer program. It should also provide insight into the hotly debated topic of the degree to which skillful poker play consists of an understanding of numeric probability and how much is sheer bluff. Imperfect information characterizes the world in which we live, and the development of a computer capable of functioning at a high level at a poker table would be a major breakthrough for the artificial intelligence community, with implications far beyond the gambling world. My money’s on the Unabomber.® Burke Hansen, attorney at large, heads a San Francisco law office GeekList: Free Computer Version of Board and Card Games with Artificial Intelligence Computer Oppone
boardgamegeek.com/geeklist/8323
The Chronicle: Wired Campus Blog: Team Studies Artificial Intelligence With Poker-Playing Computer
chronicle.com/wiredcampus/article/1492/team-studie...
O'Reilly Network Safari Bookshelf - AI Techniques for Game Programming
safari.oreilly.com/193184198X?a=102682
ZDNet India > News > software > Algorithm helps computers beat humans at Go
www.zdnetindia.com/news/software/stories/172306.ht...
Computer cashes in big at Texas Hold 'Em tourney
arstechnica.com/news.ars/post/20080713-computer-ca...
O'Reilly Network Safari Bookshelf - AI Game Development: Synthetic Creatures with Learning and React
safari.oreilly.com/1592730043?a=102682
Checkers, Solved! By Suhas SreedharMake no mistakes and it's a draw, says computer scientist
19 July 2007—You might have to find a new game to play on your computer or cellphone soon, because checkers will only frustrate you. Even if you play it perfectly—without any mistakes, choosing the best strategy in each situation—you’ll have absolutely no chance of winning. One of the most popular board games in history, checkers has been definitively proved to end in a draw when played perfectly by both sides, according to Jonathan Schaeffer, professor of computer science, and his colleagues at the University of Alberta, Edmonton, Canada. They published their proof today on the Web site of the journal Science . It took Schaeffer nearly 16 years to complete his checkers odyssey. He began by inventing Chinook, the world’s first computer program to win a world championship against a human in checkers, or any other game, for that matter. The program, created in 1989, draws from a library of stored opening moves that had been played by grandmasters, an endgame database that traces backward from the final results of games (wins, losses, or draws), and a deep-search algorithm that makes decisions during play based on evaluating all possible outcomes a certain number of moves ahead.
In 1990, Chinook won the right to compete in the World Championship by coming in second place at the United States National Open Checkers Championship to Marion Tinsley, considered to be the greatest checkers player who ever lived. Chinook was all set to face Tinsley in the World Championship, but the American Checker Federation (ACF) and the English Draughts Association (EDA) had not sanctioned the match. Tinsley, wanting to face the program, resigned his title in protest. The ACF and EDA soon created the Man vs. Machine World Championship to accommodate the match. Chinook lost to Tinsley, with two wins to Tinsley’s four. In 1994, Chinook was retooled with more information, including previous strategies used by Tinsley, and faced off against the grandmaster in a rematch. All six games played resulted in draws. Tinsley, who was suffering from pancreatic cancer, withdrew from the tournament and died seven months later. Chinook was declared the Man vs. Machine World Champion and went on to defend its title against grandmaster Don Lafferty by winning one match, losing one, and then drawing 18. After that defense, Schaeffer retired Chinook from playing and put it to use in helping to solve checkers. Solving checkers, a game played on a board with 64 squares using 12 black and 12 white or red pieces, was a daunting task. There are 500 billion billion (5 x 1020) possible situations that could arise while playing. Strongly solving the game or computing all of these possible positions would have taken decades, says Schaeffer. Instead, he implemented a two-pronged search technique that concluded with the game being a draw by examining only a fraction of all the possible game scenarios. 
First, he constructed databases of endgames, building backward from all the possible wins, losses, or draws that checkers could conclude with. A so-called backward-searching algorithm built the path of situations that would have led to these endgames all the way to the point where there were 10 game pieces on the board. The result is a database of 39 trillion positions compressed using a homebrew algorithm into an average of 237 gigabytes for an average of 154 positions per byte of data. When Schaeffer first developed these databases, it took him seven years—from 1989 to 1996—to build backward from the endgame to the point where there were eight pieces on the board. Deeming that there wasn’t enough horsepower or RAM available to compute any further, in 1997 he suspended the project. When he resumed the project in 2001, computing power had so improved that the endgame databases that had taken years to compute were redone in the course of a month. From there the eight-piece scenarios were extended back to 10-piece scenarios. The next step was to use a forward-search technique, such as the ones chess software typically rely on to figure out how to get to those 10-piece situations from the beginning of the game, when all 24 pieces are on the board. This forward search, however, was not performed in the way Chinook or IBM’s Deep Blue—the chess-playing computer that defeated world champion Garry Kasparov 10 years ago—would have done it. Those programs used deep-search algorithms to make their next moves by analyzing all possible situations that are one-move deep, then all possible situations that are two-moves deep, and so on. Instead, Schaeffer and his colleagues used a technique called “best first” to prioritize searching various positions and lines of play. At a given position in the game there are several possible moves that can be made. Instead of exploring all of these moves to their final outcomes using deep search, Schaeffer's team used Chinook to provide a measure of what the strongest line of play would be—what would most likely result in a win in the fewest moves. This line of play was evaluated first. If it did result in a win, then there was no need to search any other parallel lines of play, because the entire line was already known to result in a strong win. A win in such an instance is not a characteristic of perfect play by both sides; perfect play means that each side will try to win in as few moves as possible, or delay losing in as many moves as possible. Since a win was achieved so quickly, it means the losing side made a mistake and did not play perfectly. Entire lines of play branching from various positions were eliminated this way, vastly reducing the number of lines that had to be deeply explored. By applying such a technique, Schaeffer’s team was able to solve checkers using the least amount of effort. Of the 5 x 1020 possible positions, Schaeffer needed to evaluate only 1014 to prove that checkers, played perfectly, results in a draw. Players had suspected for a while that checkers would result in a draw, says Schaeffer. Human players draw so frequently when playing that since 1934 championship tournaments require the first three half-moves (black-white-black) to be done randomly from a list of accepted openings in order to reduce the number of draws. Schaeffer’s proof solved checkers for 19 different openings, all of which end in draws. There are 300 total tournament openings, but many of these were determined to either be mirrors of other positions or altogether irrelevant to the proof because they lead to positions common to other openings. Solving checkers has taken a big monkey off Schaefer’s back. The fact that the game wasn’t solved for every possible position and then tucked away in a database doesn’t seem to bother him. “Well, the checkers players would love it, because [then] you’ve got this oracle that can tell them everything—answer every single unanswered question in the game of checkers,” he says. “But first of all, I don’t have the patience to do it. And second of all, I don’t have the technology to do it.” Even with the best data-compression techniques, Schaeffer says, the amount of storage required to solve all possible positions of checkers would exceed even the capacity of the world’s biggest supercomputers with tens of petabytes (1015) of storage by an order of magnitude. That puts it—at the earliest—at least a decade away. But there is little motivation for Schaeffer to pursue such a quest. He has solved checkers and has painstakingly verified that none of the data were corrupt or inaccurate. Vasik Rajlich, an international chess master [profiled in “Dream Jobs,” IEEE Spectrum, February 2007 described the accomplishment as the latest in a line of games that have been solved computationally. “Every now and then some game is solved,” he says. “And now we can ‘check this box’ for a rather major game.” So far researchers have solved Connect Four, Qubic, Go-Moku, Nine Men’s Morris, and Awari. As for the question of solving a game like chess, which people suspect will also result in a draw, the amount of data is even more monstrous. The number of positions in checkers is thought to be roughly the square root of the number of positions in chess. That’s somewhere in the order of 1040 to 1050 positions. Schaeffer says that even with the two-pronged technique he used in solving checkers, a breakthrough such as quantum computing would be needed to even attempt to solve chess. But he isn’t quick to rule out the possibility. “The one thing I’ve learned in all of this is to never underestimate the advances in technology,” he says. Slashdot | Researchers Create an Automatic Backup Band for Singers
tech.slashdot.org/article.pl?sid=08/04/07/1550226&... Researchers Create an Automatic Backup Band for Singers
Posted by
ScuttleMonkey
on Monday April 07, @12:47PM
from the william-hungs-of-the-world-unite dept. 
Researchers at Microsoft Labs are hoping to allow untrained singers to have their own automatic backup band in the near future. A new piece of software, "MySong", promises to take a sung melody and using a probability computation algorithm, generate an appropriate chord accompaniment. There is also a video of the process on the Microsoft Labs website. "'The idea is to let a creative but musically untrained individual get a taste of song writing and music creation,' Morris told New Scientist. 'There was nothing out there that could take a sung vocal melody as an input and then generate appropriate chords to accompany it. [...] Since people rarely sing at precise frequencies, MySong compares a sung melody to the 12 standard musical notes. It then feeds an approximate sequence of notes to the system's chord probability computation algorithm. This algorithm has been trained, through analysis of 300 rock, pop, country and jazz songs, to recognize fragments of melody and chords that work well together, as well as chords that complement each another.'"
Math Trek: Tetris Is Hard, Science News Online, Oct. 26, 2002
www.sciencenews.org/articles/20021026/mathtrek.asp Tetris Is HardIvars Peterson As many computer- and video-game players have long known, the insanely addictive, immensely popular game of Tetris is tough. You can't really win; you merely try your best to improve upon previous results.
The game was invented in 1985 by mathematician Alexey Pajitnov, then a computer engineer at the Academy of Science's Computer Center in Moscow. The game board is a rectangular grid of squares, initially occupied by a given configuration of filled squares. The player is given, one by one, a sequence of what are called p tetrominoes. A tetromino is a set of four squares (or blocks) arranged into a larger square, a "T," an "L," or some other configuration. Each piece starts in the middle of the top row of the game board and falls downward at a constant speed. As it falls, the player can rotate the piece or slide it sideways. Pieces travel downward as far as they can go, then stop, permanently fixed in place. As soon as one piece reaches its resting spot, the next one begins its descent. If, when a piece comes to rest, all the squares in an entire row of the grid are filled, that row is cleared. All higher rows drop down one row. A player loses when a new piece is blocked by filled grid squares and cannot descend at all. The player's goal is to maximize his or her score, which increases as pieces are placed and rows cleared. Now, Erik D. Demaine, Susan Hohenberger, and David Liben-Nowell of the Massachusetts Institute of Technology's Laboratory for Computer Science have analyzed Tetris from a computational perspective, focusing on the computer resources required to play the game successfully. In effect, they determined the relative running time or amount of memory a computer would require to play the game in its most demanding form. The researchers initially focused on a version of Tetris in which the player knows the identity and order of all the pieces that will be presented. They also allowed the player an arbitrary number of shifts and rotations before a piece dropped into place. We studied this version "because its hardness captures much of the difficulty of playing Tetris," they remarked. "Intuitively, it is only easier to play Tetris with complete knowledge of the future, so the difficulty of playing the offline version suggests the difficulty of playing the online version." To get a measure of the computational complexity of Tetris, Demaine and his coworkers considered a generalized version of the game—one in which the game board grid could be any number of squares wide and high. In this context, the computer scientists discovered that maximizing the number of rows cleared while playing the given sequence of pieces belongs to a category of problems described as NP-complete. An NP problem is one for which it is relatively easy to check whether a given answer is correct, but may require an impossibly long time to solve by any direct procedure. Interestingly, the computer game Minesweeper also belongs to the NP-complete category. It turns out that other goals of Tetris, such as maximizing the number of pieces placed before a loss occurs or minimizing the height of the highest filled grid square, also belong to the NP-complete category. So, even if you know all the pieces in advance and can take as long as you want for each move, the game is still challenging. Demaine and his collaborators also looked at other variants of Tetris. They evaluated, for example, the effect of restrictions on rotations, limitations on player agility, and the use of smaller sets of different tetrominoes. What about the computational complexity of the actual game, where the sequence of pieces is generated randomly and the pieces can fall very quickly? That's still an open question.
References: Demaine, E.D., S. Hohenberger, and D. Liben-Nowell. Preprint. Tetris is hard, even to approximate. Abstract available at http://xxx.lanl.gov/abs/cs.CC/0210020. Kaye, R. 2000. Minesweeper is NP-complete. Mathematical Intelligencer 22(No. 2):9-15. See also http://www.mat.bham.ac.uk/R.W.Kaye/minesw/minesw.htm. Peterson, I. 2002. Logic in the blocks. Science News 162(Aug. 17):106-108. Available at http://www.sciencenews.org/20020817/bob10.asp. ______. 1999. Minesweeper logic. Science News Online (May 1). Available at http://www.sciencenews.org/sn_arc99/5_1_99/mathland.htm. The official Tetris Web site can be found at http://www.tetris.com/index_front.html. Erik Demaine has a Web page at http://theory.lcs.mit.edu/~edemaine/. For information about Minesweeper as an NP-complete problem, see http://www.claymath.org/prizeproblems/milliondollarminesweeper.htm. ********** A Decade After Kasparov's Defeat, Deep Blue Coder Relives Victory -
www.wired.com/science/discoveries/news/2007/05/mur... A Decade After Kasparov's Defeat, Deep Blue Coder Relives Victory
Robert Andrews
05.11.07 | 2:00 AM
Murray Campbell
On this day 10 years ago, the human race got an inferiority complex. A computer, Deep Blue, beat Russian Garry Kasparov, the greatest chess player on the planet, and mankind’s place in the order of things was reshuffled. Blame IBM. Deep Blue was just the latest in a line of three supercomputers developed by Big Blue’s research scientists over the decade before its triumph in New York on May 11, 1997. The first, Deep Thought, emerged in 1988 from the drawing board of Carnegie Mellon grad student Feng-hsiung Hsu, who was hired by the computer maker a year later with the express intention of usurping the primacy of human logic. Upon joining IBM, Hsu recruited several programmers, including his computer science classmate, Canadian Murray Campbell, with whom he had prototyped an early chess-playing computer named Chiptest several years earlier. Kasparov, who was the youngest-ever world chess champion, trounced the first Deep Thought in a 1989 matchup. A much-improved machine, benefiting from IBM’s considerable resources, proved stiffer competition in a February 1996 rematch. It was the third incarnation, dubbed Deep Blue, that finally knocked the master from his perch 15 months later. The match immediately became an iconic symbol of the advances made in artificial intelligence and supercomputing. Kasparov has since retired, like Deep Blue, which now resides in a museum. He has become a vocal advocate for democracy in today’s Russia. But Campbell, who moved the pieces on Deep Blue’s command, is still at IBM in New York state; he told Wired News he would mark the anniversary by watching videos of the contest. Wired News: What is the state of supercomputer-versus-human matchups? How are we humans doing? Murray Campbell: Not so well! The current world champion, Vladimir Kramnik from Russia, lost a match to a PC program in November, 4-2. If you look at the supercomputer that Deep Blue ran on, I think a present-day Cell processor has as much processing power as that entire system did in 1997. WN: What’s going to be the next move in supercomputing over the next 10 years? Campbell: It's almost the end of the story for chess in the sense that matches between chess machines and grand masters are becoming less interesting because it's so difficult for the human grand masters to compete successfully. They're even taking relatively dramatic steps like giving handicaps to computers, making them play the game with a pawn less or playing the game with less time. We're past the stage where there's a debate about who's better -- machines or grand masters -- and we're just looking for interesting ways to make the competition fairer. WN: So we can live in harmony? Campbell: Absolutely. Very nicely. WN: Why do you think the match captured the public imagination to such a great degree? Campbell: Not everybody plays but everybody knows that chess is a very difficult game -- you can devote your whole life to playing chess and still have room for improvement. So it's understood that chess is a game which requires intelligence and so for a machine to play at the level of the world champion is a sign that computers have progressed further than maybe some people had thought they had. In one sense, it's just one milestone among many along the way, but it's one that the general public could understand more easily. WN: Some conspiracies emerged that say the computer was being controlled by human grand masters behind the scenes. Your response? Campbell: In my mind, they're ridiculous. One of the things about conspiracy theories is that, really, you should present some evidence rather than just make a theory up. In the 10 years since the match, I have yet to see any evidence of cheating. WN: What are Deep Blue’s roots, and on what technological principles did its forebears operate? Campbell: Claude Shannon, the famous computer scientist and mathematician proposed that chess was a grand challenge for these new things called computers -- if you could get a computer to play chess at the world champion level, you had done something really special. There was a turning point in the '70s when it was realized that, if you let computers do what they do best -- that is, search through as many possibilities as they can as quickly as they can -- and stop the pretense of trying to emulate how humans play, you actually got better performance. And so, from that day on, computers, including Deep Blue, tended to be focused on searching through as many possible chess moves as they could in the amount of time that was available for a computation. WN: How did the notion come about of a contest with Kasparov? It was obviously going to be a high-profile match. Campbell: I don't think it was obvious at all. Actually, at the time, in 1989, it was a fairly low-key event. There were just two games between Deep Thought and Kasparov and I think many people were surprised at the level of interest that that match drew. WN: Kasparov easily beat Deep Thought 2-0. Was the machine a sore loser? Campbell: No, it took it pretty well. WN: The buildup to the first Deep Blue game -- were you guys at all nervous? Campbell: Yeah, it was a bit of a nerve-wracking time, mainly because the hardware that Deep Blue ran only got working a few weeks before the match. So there was frantic testing before the first match. Pretty much it was right down to the wire, so we obviously were quite pleased when it won the first game of that match, even though it lost the match. WN: Deep Blue got some software modifications and hardware upgrades for its next match that helped it analyze 200,000,000 moves per second. That's some halftime pep talk. Campbell: That was one of the factors we ran on a larger IBM supercomputer, but we also redesigned the chess-specific hardware so it would run more chess patterns. It just understood more about chess and was able to play better because of it. We told it what it was doing wrong, in a sense, and hoped that would be fixed. WN: You took input from several chess experts in developing Deep Blue. Do you think that put the machine at an unfair advantage over Kasparov? Campbell: No. We obviously were going to improve the system and work with grand masters to help us do that. Kasparov, of course, had a team of assistants to help him as well. WN: By 1997, the machine actually played in a way that was quite human, even making a mistake at one point, while Kasparov was forced to change his usually liberal game for a more robotic style. What on earth was going on? Campbell: Conventional wisdom then, and even now, is that you don't want to go into complications against a strong chess computer because you will lose -- you'll get out-calculated. And so this was Kasparov's attempt to keep the game more strategic, more under control. WN: What did it feel like to win finally? Campbell: It was obviously a highlight of my career. I was a chess player growing up and so to be sitting across the chess board from Kasparov was a thrill in itself. WN: Kasparov's request for a rematch was turned down. Why do you think that was? It looked like IBM had got good publicity but took the ball away and didn't want to play anymore. Campbell: I believe we had accomplished what we were trying to do for so many years and it was time to move on to other problems. Garry Kasparov has now retired from chess and Deep Blue is in a museum so that's not going to happen, but matches between other computers and the current world champion have taken place as recently as last year. WN: What did the contest do for IBM and you personally? Campbell: IBM was an also-ran in supercomputing, it was not one of the top two in the field. Nowadays, IBM has nearly half of the top 500 supercomputers in the world, it dominates this market now and it certainly has the fastest supercomputer in the world as well. Campbell: For me personally, I've had a chance over the years since then to become more of a user of supercomputers than a designer of supercomputers. I've done a lot of work in data analysis projects in financial and public health and the petroleum industry so I've come to appreciate the things that supercomputers can do in solving complex problems in these industries. Comment on this story. 
Theory of Everything - Algorithmic Theory of Everything - Computer Universe - Computable Universe -
www.idsia.ch/~juergen/computeruniverse.html Computable Universes & Digital Physics: Is our universe just the output of a deterministic computer program? As a consequence of Moore's law, each decade computers are getting roughly 1000 times faster by cost. Apply Moore's law to the video game business. As the virtual worlds get more convincing many people will spend more time in them. Soon most universes will be virtual, only one (the original) will be real. Then many will be led to suspect the real one is a simulation as well. Some are already suspecting this today. Then the simplest explanation of our universe
is the simplest program that computes it. In 1997
Schmidhuber
pointed out [1] that the simplest such program
actually computes all possible universes with all
types of physical constants and laws, not just ours.
His essay also talks about universes
simulated within parent universes in nested fashion,
and about universal complexity-based measures on
possible universes.
Prior measure problem: if every possible future exists, how can we predict anything? Unfortunately, knowledge about the program that computes all computable universes is not yet sufficient to make good predictions about the future of our own particular universe. Some of our possible futures obviously are more likely than others. For example, tomorrow the sun will probably shine in the Sahara desert. To predict according to Bayes rule, what we need is a prior probability distribution or measure on the possible futures. Which one is the right one? It is not the uniform one: If all futures were equally likely then our world might as well dissolve right now. But it does not. Some think the famous anthropic principle (AP) might help us here. But it does neither, as will be seen next. Anthropic Principle does not help The anthropic principle (AP) essentially just says that the conditional probability of finding oneself in a universe compatible with one's existence will always remain 1. AP by itself does not have any additional predictive power. For example, it does not predict that tomorrow the sun will shine in the Sahara, or that gravity will work in quite the same way - neither rain in the Sahara nor certain changes of gravity would destroy us, and thus would be allowed by AP. To make nontrivial predictions about the future we need more than AP - see below!
Man still top dog at pokerMan vs Machine reduxPublished Friday 27th July 2007 08:57 GMT Mobile computing: Opportunities and risk - Free whitepaper
Wireless Email Solutions - Free whitepaper');}
The Association for the Advancement of Artificial Intelligence (AAAI) wrapped up its Man vs Machine Poker Challenge yesterday, and, as we had suspected, the human element won out, although the margin of victory was not overly large. Poker is a tricky game to teach a machine. Computers are inherently deterministic and uncreative, which is why a game of "imperfect information", such as poker, is challenging to reduce to mathematical formulae sufficiently adaptable to handle chaotic real life scenarios yet reducible to stable computer code. Although a computer can calculate to perfection the odds of receiving a particular hand of cards, which is certainly a big part of the game, it cannot form judgments about whether a certain player is lying or not - which many poker devotees argue is even more important. It can, however, approximate such judgments by storing an internal record of how a particular player plays certain kinds of hands to search for tendencies. This is where most of the AI work on computerized poker games goes: developing algorithms that track opponents' betting habits. Games such as chess or checkers can be replicated well on a computer since such playing style judgments, though occasionally helpful in chess, for example, sheer number crunching can map out an overwhelming number of possible outcomes. The chess board is a closed, determinative place, easily modeled in code. The Man vs Machine Poker Challenge sought to eliminate as much of the luck as possible by playing a poker variant known as duplicate poker, in which one teammate plays the same hand as the other teammate's opponent, and vice versa. By eliminating luck, duplicate poker reduces the game to its strategic fundamentals. Phil "the Unabomber" Laak and Ali "the Prince" Eslami, professional players known for their mathematical dexterity, edged out the University of Alberta's computer poker program Polaris by about $400 over the course of the two day low limit match. The more interesting work is yet to come, however; playing Texas Hold'em against one person, where only two cards are unknown, and one player's tendencies are to be studied, has been challenging enough for those keen on developing championship caliber poker software. Developing software that deals with multiple lying individuals and ten or more unknown cards will be exponentially more difficult. Duplicate poker may be helpful in measuring scientifically how different individuals react to comparable situations, but the variant itself is something of a red herring, at least when it comes to developing strategic poker software. For one thing, by eliminating luck, the variant rewards the mathematically gifted, which plays to the computer's strength. The variant also eliminates one of the most powerful psychological components of the game itself - the winning (or losing) streak. Such a psychological component may not exist in the program itself, but when it comes to the analyzing the psychological habits of an opponent, buffeted as that player may be by the randomness of the cards, the complexity of the game magnifies with the good or bad fortune that befalls that player. Players may change their strategies - not to mention the amount of the wager - completely based on how confident or wealthy they feel at that time, and what cards have been good to them. Gamblers are a superstitious lot, and superstition has no rhyme or reason. The machine has a long way to go. ® Burke Hansen, attorney at large, heads a San Francisco law office Microsoft: Please Make MySong Into a Web App - ReadWriteWeb
www.readwriteweb.com/archives/microsoft_mysong.php Microsoft: Please Make MySong Into a Web App
Written by Josh Catone / April 7, 2008 4:02 PM
/ 3 Comments
MySong is a project from Microsoft researchers Dan Morris, Sumit Basu, and Washington State grad student Ian Simon. The program is made to "give many folks who would never even taste songwriting a great opportunity to just get a glimpse of music creation," according to the project's web site. It is "more than good enough to make a cute birthday song for Mom or a Valentine's Day song for your significant other." Sounds like a great premise for a web app to us! The program works by identifying the 12 standard musical notes in a sung melody, and then feeding those notes into an algorithm that has been trained by listening to 300 songs in varying genres and learning how to identify chords and melody fragments that work well together. The result is a series of musical accompaniments that users can adjust via sliders for "happy factor" and "jazz factor." "I suspect musicians will argue that this is another step towards homogenized elevator music for all," Peter Bentley, a computer scientist at University College London, told New Scientist. "But I see a big market for this, whether it's liked by musicians or not." We agree, and we think the web is the perfect place to find that market. Last year we reviewed a startup backed by Pete Townsend of "The Who," Method Music, that created personalized theme music based on user input (specifically: a voice sample, a picture, and a recording of a rhythm). The results were... not unimpressive, but neither were they very compelling. MySong, on the other hand, produces some very impressive (to my musically untrained ear) output and has a much more compelling set of use cases. Though Microsoft hasn't decided how or if to market the MySong technology, Morris told the New Scientist that it wasn't very computationally demanding. "It could even run on a cellphone," he said. Check out this sample of voice-only input, the MySong output, and a full musical arrangement after being fed through Band-in-a-Box. You can draw your own conclusions, but we were very impressed and hope that someone brings the MySong technology to a web browser in the future. We have a feeling this thing would kill as a Facebook or OpenSocial application. There are many more samples on the MySong page. 
Computers versus Common Sense
for
The download is starting. Please Save the video file.
If the download does not start automatically, right-click this link and choose "Save As".
Questions? Check the Download FAQ
Copy and paste the HTML below to embed this video onto a web page.
Google TechTalks
May 30, 2006
Douglas Lenat Dr. Douglas Lenat is the President and CEO of Cycorp. Since 1984, he and his team have ... all » been constructing, experimenting with, and applying a broad real world knowledge base and reasoning engine, collectively "Cyc". Dr. Lenat was a professor of computer science at Carnegie-Mellon University and at Stanford University. His interest and experience in national security has led him to regularly consult for several U.S. agencies and the White House. ABSTRACT It's way past 2001 now, where the heck is HAL? For several decades now we've had high hopes for computers amplifying our mental abilities not just giving us access to relevant stored information, but answering our complex, contextual questions. Even applications like human-level unrestricted speech understanding continue to dangle close but just out of reach. What's been holding AI up? The short answer is that while computers make fine idiot savants, they lack common sense: the millions of pieces of general knowledge we all share, and fall back on as needed, to cope with the rough edges of the real world. I will talk about how that situation is changing, finally, and what the timetable -- and the path -- realistically are on achieving Artificial Intelligence.
An introduction to rules engines [printer-friendly] | Reg Developer
www.regdeveloper.co.uk/2006/08/03/rules_engines_pa...
Tech Report: HPL-2006-20R1: Tackling Concept Drift by
www.hpl.hp.com/techreports/2006/HPL-2006-20R1.html...
» Can computers have an opinion? | Emerging Technology Trends | ZDNet.com
blogs.zdnet.com/emergingtech/?p=364
Slashdot | Open Source Robot for Household Tasks
hardware.slashdot.org/article.pl?sid=08/03/07/0122... Open Source Robot for Household Tasks
Posted by
Soulskill
on Thursday March 06, @11:58PM
from the make-them-do-the-dishes-when-they-beat-you-at-chess dept. 
bednarz brings us a NetworkWorld story about the development of a robot through an open source project. The objective of the project is to "take robotics from research into homes." Quoting:
"One of its immediate goals is to build 10 robots and make them available to university researchers as a common platform that can be tinkered with and improved. Willow Garage will also supply 'an open-source code base integrated from the best open-source robotics software available,' President and CEO Steve Cousins said. In Cousins' video presentation, the first version of the robot could be seen vacuuming, picking up toys off the floor of a living room, taking dishes out of a dishwasher, and most importantly of all, using a bottle opener to crack open a cold, refreshing brew."
www.technologyreview.com/printer_friendly_article.aspx?id=19068
www.technologyreview.com/printer_friendly_article.... Thursday, July 19, 2007
Robotic Insect Takes Off for the First Time
Researchers at Harvard have created a robotic fly that could one day be used for covert surveillance and detecting toxic chemicals.
By Rachel Ross
A life-size, robotic fly has taken flight at Harvard University. Weighing only 60 milligrams, with a wingspan of three centimeters, the tiny robot's movements are modeled on those of a real fly. While much work remains to be done on the mechanical insect, the researchers say that such small flying machines could one day be used as spies, or for detecting harmful chemicals. "Nature makes the world's best fliers," says Robert Wood, leader of Harvard's robotic-fly project and a professor at the university's school of engineering and applied sciences. The U.S. Defense Advanced Research Projects Agency is funding Wood's research in the hope that it will lead to stealth surveillance robots for the battlefield and urban environments. The robot's small size and fly-like appearance are critical to such missions. "You probably wouldn't notice a fly in the room, but you certainly would notice a hawk," Wood says. Recreating a fly's efficient movements in a robot roughly the size of the real insect was difficult, however, because existing manufacturing processes couldn't be used to make the sturdy, lightweight parts required. The motors, bearings, and joints typically used for large-scale robots wouldn't work for something the size of a fly. "Simply scaling down existing macro-scale techniques will not come close to the performance that we need," Wood says. Some extremely small parts can be made using the processes for creating microelectromechanical systems. But such processes require a lot of time and money. Wood and his colleagues at the University of California, Berkeley, needed a cheap, rapid fabrication process so they could easily produce different iterations of their designs. Ultimately, the team developed its own fabrication process. Using laser micromachining, researchers cut thin sheets of carbon fiber into two-dimensional patterns that are accurate to a couple of micrometers. Sheets of polymer are cut using the same process. By carefully arranging the sheets of carbon fiber and polymer, the researchers are able to create functional parts. For example, to create a flexure joint, the researchers arrange two tiny pieces of carbon composite and leave a gap in between. They then add a sheet of polymer perpendicularly across the two carbon pieces, like a tabletop on two short legs. Two new pieces of carbon fiber are placed at either end of the polymer, as a final top layer. Once all the pieces are cured together, the resulting part resembles the letter H: the center is flexible but the sides are rigid. By fitting many little carbon-polymer pieces together, the researchers are able to create rather complicated parts that can bend and rotate precisely as required. To make parts that will move in response to electrical signals, the researchers incorporate electroactive polymers, which change shape when exposed to voltage. The entire fabrication process will be outlined in a paper appearing in an upcoming edition of the Journal of Mechanical Design. After more than seven years of work studying flight dynamics and improving various parts, Wood's fly finally took off this spring. "When I got the fly to take off, I was literally jumping up and down in the lab," he says. Other researchers have built robots that mimic insects, but this is the first two-winged robot built on such a small scale that can take off using the same motions as a real fly. The dynamics of such flight are very complicated and have been studied for years by researchers such as Ron Fearing, Wood's former PhD advisor at the University of California, Berkeley. Fearing, who is building his own robotic insects, says that he was very impressed with the fact that Wood's insect can fly: "It is certainly a major breakthrough." But Fearing says that it is the first of many challenges in building a practical fly. At the moment, Wood's fly is limited by a tether that keeps it moving in a straight, upward direction. The researchers are currently working on a flight controller so that the robot can move in different directions. The researchers are also working on an onboard power source. (At the moment, the robotic fly is powered externally.) Wood says that a scaled-down lithium-polymer battery would provide less than five minutes of flying time. Tiny, lightweight sensors need to be integrated as well. Chemical sensors could be used, for example, to detect toxic substances in hazardous areas so that people can go into the area with the appropriate safety gear. Wood and his colleagues will also need to develop software routines for the fly so that it will be able to avoid obstacles. Still, Wood is proud to have reached a major project milestone: flight. "It's quite a major thing," he says. "A lot of people thought it would never be able to take off." Copyright Technology Review 2007. iRobot inks deal for laser-radar droidvision sensors | The Register
www.theregister.co.uk/2008/01/29/irobot_ladar_lase... iRobot inks deal for laser-radar droidvision sensorsNow the machines can follow you into the buildingPublished Tuesday 29th January 2008 13:27 GMT
iRobot, provider of ground warbots to the US forces and purveyor of domestic droids to the comfortably off consumer, has struck a deal allowing it to use laser-scanning technology in its future designs. Reports have it that the kit could be in use in 2009; though this would be with the military, not on the company's famous line of autonomous "Roomba" floor-cleaners. According to the Christian Science Monitor, the laser-vision system - also known as "ladar", as it is effectively a laser-light version of radar rather than a camera - is provided by Advanced Scientific Concepts of California. iRobot has also bought shares in ASC, as well as inking the exclusive marketing and technology deal. The new gear is intended to solve one of the most difficult problems in autonomous ground-mobile machines - that of generating a useable 3D map of what lies ahead. Humans can do a good job of this using stereoscopic 2D vision, but thus far software has struggled to interpret ordinary camera imagery in a way that robots can use. This has tended to mean that robots which rely on cameras as sensors need to be operated remotely by humans; which in turn calls for a high-bandwidth, low-latency datalink and expensive dedicated personnel. One solution is to use radar, which maps objects in relation to the detector, but even millimetre-wave radar becomes hard to use in the close, cluttered environments which robots must navigate. The ASC systems will use brief pulses of laser light instead of radio or microwaves, thus perhaps allowing autonomous operations inside buildings - an attractive option in iRobot's main markets. The 5-nanosecond pulses will be eye-safe, apparently. Ladar detection has already been used in this way, perhaps most famously in the DARPA Grand Challenge robot-car competitions where contenders face similar problems to iRobot. However the Grand Challenge vehicle ladars were mostly unsuitable for production, and for the indoor applications iRobot has in mind. ASC's "laser flash" technology, originally developed for aerial mapping, is nearer being ready to go. iRobot believes it might have ladar on military combat droids in 12 to 18 months, and thereafter on its household machines at some unspecified date.® 14 comments posted — Post a new comment
Furby redux? Ars Technica reviews the PleoBy Jacqui Cheng | Published: February 12, 2008 - 11:31PM CT I, for one, welcome our new robot dinosaur overlordsSince the Pleo's introduction to the world last year, the little robot dinosaur has been hailed as one of the most innovative toys of 2007. At first blush, it's not hard to see why. Pleo (referred to as Pleo, not "the" Pleo, just like Steve Jobs and Apple call the iPhone "iPhone") is adorable, contains sensors all over its body, and is programmed with a wide range of reactive emotions. Ugobe, Pleo's creator, describes Pleo as an autonomous dinosaur. Although this promise comes from the people who also created the much-maligned Furby, the description is intriguing, as are descriptions of Pleo's capabilities from the company's web site, and the short videos circulated online that are used to hook users (owners?) on Pleo's cuddly demeanor. "Awareness glimmers as Pleo adjusts to the light. His limbs try a tentative stretch. The world is a flood of sensations. He'll notice you as soon as he's ready. Watch… Wait… Nurture… Your soft words and soothing touch are just what he needs." This thing sounds a little like the successor to the now-canceled Sony Aibo—a (roughly) $2,000 robotic dog that never took off—except much more affordable at $350, and in dinosaur form. I had high hopes for our Pleo when it arrived, and decided to give it a test drive to see whether it lived up to all these claims. My Pleo's name is Herbert. Welcome to the world, HerbertHerbert arrived in a box that showcased him and all of his accessories. It's a little creepy to see him wrapped up like this at first (it kind of reminded me of the movie E.T., when they thought E.T. had died and zipped him up in a body bag). Worry not; Pleo comes to life in no time. 
Herbert came with a leaf (used for eating and playing), two batteries, a battery charger, and a USB cable (for software updates). His battery needed to be charged fully before we could start playing with him, and doing this requires you to remove the battery from his belly and plug it into the included charger. The Pleo, unfortunately, cannot charge himself, and therefore you must constantly make sure there's a battery ready to go when you're planning to play with it. More on batteries later. Do Artificial Intelligence Chatbots look like their programmers? [printer-friendly] | The Register
www.theregister.com/2006/09/29/artificial_intellig...
Sorting facts and opinions for Homeland Security
www.eurekalert.org/pub_releases/2006-09/cuns-sfa09...
Slashdot | Natural Language Processing for State Security
it.slashdot.org/article.pl?sid=06/09/25/0111231&fr...
www.technologyreview.com/printer_friendly_article.aspx?id=17518
www.technologyreview.com/printer_friendly_article....
AI prof: The robot terrorists are coming! Aiee! | The Register
www.theregister.co.uk/2008/02/27/sharkey_robot_ter... AI prof: The robot terrorists are coming! Aiee!Noel Sharkey issues another robopocalypse warningPublished Wednesday 27th February 2008 11:17 GMT
Noted Brit media prof Noel Sharkey - perhaps most famous as a judge on TV's Robot Wars - has warned (again) of the coming danger to humanity posed by killer robots. But this time, rather than military deathdroids or homicidal mechanoid nurses (see below), Sharkey is flagging up the threat from killbots directed by sinister militants or guerillas. Yes, you read that right: the ATTACK OF THE ROBOT TERRORISTS is imminent! Flee, oh flee for your lives while you still can! According to a Reuters report, Sharkey - whose day job is Professor of Artificial Intelligence and Robotics at Sheffield Uni - will deliver his latest robopocalypse warning to a military thinktank tonight. Pointing to the burgeoning droid armies of America and its allies, the prof reckons that enemies of freedom and democracy will soon get in on the act - perhaps triggering a fearful orgy of mechanised slaughter which could, unchecked, wipe out humanity. Or maybe see the few survivors scurrying like rats in underground tunnel networks or something. "How long is it going to be before the terrorists get in on the act?" asks Sharkey. "With the current prices of robot construction falling dramatically and the availability of ready-made components for the amateur market, it wouldn't require a lot of skill to make autonomous robot weapons." Apparently the prof - who boasts chartered-engineer status along with his Equity card and sheaves of psych and biobotics credentials - reckons that a basic killbot could be built for £250. He's thinking here of your everyday GPS toy-aircraft DIY cruise missile style caper. Previously, Sharkey has warned of "a robot arms race that will be difficult to stop... I can imagine a little girl being zapped because she points her ice cream at a robot to share". And that's not all. On his webpage, the roboticist says: This has become a passion for me. There is a cultural mythology about robots fed by media, goverments and scientists alike. The thinking robot is still only a fairytale. Behind the zoomorphic dream of robot companions and helpers there are some real dangers... the rise of robot elderly carers, child minders, nurses, soldiers and police... mobile robot surveillance. These are not super-intelligent robots. They are dumb automatic machines and we must decide what we want from them before we de-humanise ourselves further. Other previous Sharkey quotes: "It would be great if all the military were robots and they could fight each other" and "Imagine the miners' strike with robots armed with water cannon". He has also, apparently, predicted the imminent arrival of vibrating eroto-droid concubines - which could presumably be subverted by terrorist hackers in hilarious Austin Powers style. Holy crap. Little girls blown away by soulless droid childminder/soldiers for waving ice creams? £250 terrorist sat nav kamikaze deathdrones homing in remorselessly on their helpless targets laden with CARGOES OF DEATH? Sensible decision-making by human troops replaced by a frenzied automatic bloodbath? Or maybe - just maybe - a case of inky hacks and media profs in a robot-headline feedback loop? You decide. ® BBC NEWS | Technology | Robot unravels mystery of walking
newsvote.bbc.co.uk/mpapps/pagetools/print/news.bbc...
Robot unravels mystery of walking
Runbot is a self-learning, dynamic robot, which has been built around the theories of Nikolai Bernstein. "Getting a robot to walk like a human requires a dynamic machine," said Professor Florentin Woergoetter. Runbot is a small, biped robot which can move at speeds of more than three leg lengths per second, slightly slower than the fastest walking human. Bernstein said that animal movement was not under the total control of the brain but rather, "local circuits" did most of the command and control work. The brain was involved in the process of walking, he said, only when the understood parameters were altered, such as moving from one type of terrain to another, or dealing with uneven surfaces. The basic walking steps of Runbot, which has been built by scientists co-operating across Europe, are controlled by reflex information received by peripheral sensors on the joints and feet of the robot, as well as an accelerometer which monitors the pitch of the machine. These sensors pass data on to local neural loops - the equivalent of local circuits - which analyse the information and make adjustments to the gait of the robot in real time.
Information from sensors is constantly created by the interaction of the robot with the terrain so that Runbot can adjust its step if there is a change in the environment. As the robot takes each step, control circuits ensure that the joints are not overstretched and that the next step begins. But if the robot encounters an obstacle, or a dramatic change in the terrain, such as a slope, then the higher level functions of the robot - the learning circuitries - are used.
The latest findings of the robot research study are presented in the Public Library of Science Computational Biology journal. Four other scientists - Poramate Manoonpong, Tao Geng, Tomas Kulvicius and Bernd Porr - are also involved in the project, which has been running for the last four years. Professor Woergoetter, of the University of Gottingen, in Germany, said: "When Runbot first encounters a slope these low level control circuits 'believe' they can continue to walk up the slope without having to change anything. "But this is misguided and as a consequence the machine falls backwards. This triggers the other sensors and the highest loop we have built into Runbot - the learning circuitry - and from that experience of falling the machine knows that something needs to be changed." Dynamic process He said human walking was a dynamic process. "About half of the time during a gait cycle we are not doing anything, just falling forward. We are propelling ourselves over and over again - like releasing a spring. "In a robot, the difficulty lies in releasing the spring-like movement at the right moment in time - calculated in milliseconds - and to get the dampening right so that the robot does not fall forward and crash. "These parameters are very difficult to handle," he said.
Runbot walks in a very different way from robots like Asimo, star of the Honda TV adverts, said Prof Woergoetter. "They are kinematic walkers - they walk step by step and calculate every single angle, every millisecond. "That can be handled through engineering but it is very clumsy. No human would walk like that. All these big machines stomp around like robots - we want our robot to walk like a human." The first step in building Runbot was creating a biomechanical frame that could support passive walking patterns. Passive walkers can walk down a slope unaided, propelled by gravity and kept upright and moving through the correct mechanical physiology. Prof Woergoetter said: "Passive walking looks pretty realistic - but that's level one. On top of this we have local circuits, nested neural loops, which operate between the muscles (the joints of the robot) and the spinal cord (the spinal reflex of Runbot)." He said Runbot learned from its mistakes, much in the same way as a human baby. "Babies use a lot of their brains to train local circuits but once they are trained they are fairly autonomous. "Only when it comes to more difficult things - such as a change of terrain - that's when the brain steps in and says 'now we are moving from ice to sand and I have to change something'. "This is a good model because you are easing the load of control - if your brain had to think all the time about walking, it's doubtful you could have a conversation at the same time." Nervous system The principle was first discussed in the human nervous system by Russian physiologist Nikolai Bernstein. Prof Woergoetter said: "He said it made sense that local agents, local networks, do the basic job, but the brain exerted control whenever necessary." So using the information from its local circuits Runbot can walk on flat surfaces at speeds of more than three leg lengths per second. Prof Woergoetter said Runbot was able to learn new walking patterns after only a few trials. "If walking uphill, the gait becomes shorter, the robot's upper body weight shifts forward," he said. The key lesson from the study, he said, was that the nested loop design first proposed by Bernstein more than 70 years ago "worked and was efficient". He said the challenge was now to make Runbot bigger, more adaptive and to better anticipate situations like change of terrain.
Story from BBC NEWS: http://news.bbc.co.uk/go/pr/fr/-/2/hi/technology/6291746.stm Published: 2007/07/12 10:03:37 GMT © BBC MMVII Tech Report: HPL-2006-60R1: Pragmatic Text Mining: Minimizing
www.hpl.hp.com/techreports/2006/HPL-2006-60R1.html...
How Google translates without understanding [printer-friendly] | The Register
www.theregister.com/2007/05/15/google_translation/... How Google translates without understandingBy Bill Softky Published Tuesday 15th May 2007 00:23 GMT Column After just a couple years of practice, Google can claim to produce the best computer-generated language translations in the world - in languages their boffin creators don't even understand. Last summer, Google took top honors at a bake-off competition sponsored by the American agency NIST between machine-translation engines, besting IBM in English-Arabic and English-Chinese. The crazy part is that no one on the Google team even understands those languages.... the automatic-translation engines they constructed triumphed by sheer brute-force statistical extrapolation rather than "understanding". I spoke with Franz Och, Google's enthusiastic machine-translation guru, about this unusual new approach. Sixty years of failureEver since the the Second World War there have been two competing approaches to automatic translation: expert rules vs. statistical deciphering. Expert-rule buffs have tried to automate the grammar-school approach of diagramming sentences (using modifiers, phrases, and clauses): for example, "I visited (the house next to (the park) )." But like other optimistic software efforts, the exact rules foundered on the ambiguities of real human languages. (Think not? Try explaining this sentence: "Time flies like an arrow, but fruit flies like a banana.") The competing statistical approach began with cryptography: treat the second language as an unknown code, and use statistical cues to find a mathematical formula to decode it, like the Allies did with Hitler's famous Enigma code. While those early "decipering" efforts foundered on a lack of computing power, they have been resurrected in the "Statistical Machine Translation" approach used by Google, which eschews strict rules in favor of noticing the statistical correlations between "white house" and "casa blanca." Statistics deals with ambiguity better than rules do, it turns out. Under Google's hoodThe Google approach is a lesson in practical software development: try things and see what sticks. It has just a few major steps: 1. Google starts with lots and lots of paired-example texts, like formal documents from the United Nations, in which identical content is expertly translated into many different languages. With these documents they can discover that "white house" tends to co-occur with "casa blanca," so that the next time they have to translate a text containing "white house" they will tend to use "casa blanca" in the output. 2. They have even more untranslated text in each language, which lets them make models of "well-formed" sentence fragments (for example, preferring "white house" to "house white"). So the raw output from the first translation step can be further massaged into (statistically) nicer-sounding text. 3. Their key for improving the system - and winning competitions - is an automated performance metric, which assigns a translation quality number to each translation attempt. More on this fatally weak link below. This game needs loads of computational horsepower for learning and testing, and a software architecture which lets Google tweak code and parameters to improve upon its previous score. So given these ingredients, Google's machine-translation strategy should be familiar to any software engineer: load the statistics, translate the examples, evaluate the translations, twiddle the system parameters, and repeat. What is clearly missing from this approach is any form of "understanding". The machine has no idea that "walk" is an action using "feet," except when its statistics tell it the text strings "walk" and "feet" sometimes show up together. Nor does it know the subtle differences between "to boycott" and "not to attend." Och emphasized that the system does not even represent nouns, verbs, modifiers, or any of the grammatical building blocks we think of as language. In fact, he says, "linguists think our structures are weird" - but he demurred on actually describing them. His machine contains only statistical correlations and relationships, no more or less than "what is in the data." Each word and phrase in the source votes for various phrases in the output, and the final result is a kind of tallying of those myriad votes. Winning at chess, losing at languageThis approach is much like computerized chess: make a statistical model of the domain and optimize the hell out of it, ultimately winning by sheer computational horsepower. Like chess (but unlike vision), language is a source of pride, something both complex and uniquely human. For chess, computational optimization worked brilliantly; the best chess-playing computers, like Deep Blue, are better than the best human players. But score-based optimization won't work for language in its current form, even though it does do two really important things right The first good thing about statistical machine translation is the statistics. Human brains are statistical-inference engines, and our senses routinely make up for noisy data by interpolating and extrapolating whatever pixels or phonemes we can rely on. Statistical analysis makes better sense of more data than strict rules do, and statistical rules produce more robust outputs. So any ultimate human-quality translation engine must use statistics at its core. The other good thing is the optimization. As I've argued earlier (http://www.theregister.co.uk/2003/10/17/software_engineers_the_ultimate_brain/), the key to understanding and duplicating brain-like behavior lies in optimization, the evolutionary ratchet which lets an accumulation of small, even accidental adjustments slowly converge on a good result. Optimization doesn't need an Einstein, just the right quality metric and an army of engineers. So Och's team (and their competitors) have the overall structure right: they converted text translation into an engineering problem, and have a software architecture allowing iterative improvement. So they can improve their Black Box - but what's inside it? Och hinted at various trendy algorithms (Discriminative Learning and Expectation Maximization, I'll bet Bayesian Inference too), although our ever-vigilant chaperon from Google Communications wouldn't let him speak in detail. But so what? The optimization architecture lets you swap out this month's algorithm for a better one, so algorithms will change as performance improves. Or maybe not. The Achilles' Heel of optimization is that everything depends on the performance metric, which in this case clearly misses a lot. That's not a problem for winning contests - the NIST competition used the same "BLEU"(Bilingual Evaluation Understudy) metric as Google practiced on, so Google's dramatic win mostly proved that Google gamed the scoring system better than IBM did. But the worse the metric, the less likely the translations will make sense. The gist of the problem is that because machines don't yet understand language - that's the original problem, right? - they can't be too good at automatically evaluating language translations either. So researchers have to bootstrap the BLEU score, taking a scheme like (which merely compares the similarity of two same-language documents) and verifying that on average humans prefer reading outputs with high scores. (They compare candidate translations against gold-standard human translations) The BLEUsBut all BLEU really measures is word-by-word similarity: are the same words present in both documents, somewhere? The same word-pairs, triplets, quadruplets? In obviously extreme cases, BLEU works well; it gives a low score if the documents are completely different, and a perfect score if they're identical. But in between, it can produce some very screwy results. The most obvious problem is that paraphrases and synonyms score zero; to get any credit with , you need to produce the exact same words as the reference translation has: "Wander" doesn't get partial credit for "stroll," nor "sofa" for "couch." The complementary problem is that BLEU can give a high similarity score to nonsensical language which contains the right phrases in the wrong order. Consider first this typical, sensible output from a NIST contest: "Appeared calm when he was taken to the American plane, which will to Miami, Florida" Now here is a possible garbled output which would get the very same score: "was being led to the calm as he was would take carry him seemed quite when taken" The core problem is that word-counting scores like BLEU - the linch-pin of the whole machine-translation competitions - don't even recognize well-formed language, much less real translated meaning. (A stinging academic critique of BLEU can be found here (http://www.iccs.inf.ed.ac.uk/~osborne/papers/eacl06.pdf). A classic example of how the word-by-word translation approach fails comes from German, a language which is too "tough" for Och's team to translate yet (although Och himself is a native speaker). German's problem is its relative-to-English-tangled Wordorder; take this example from Mark Twain's essay "The Awful German Language": "But when he, upon the street, the (in-satin-and-silk-covered-now-very-unconstrained-after-the-newest-fashioned-dressed) government counselor's wife met, etc" Until computers deal with the actual language structure (the hyphens and parentheses above), they will have no hope of translating even as well as Mark Twain did here. So why are computers so much worse at language than at chess? Chess has properties that computers like: a well-defined state and well-defined rules for play. Computers do win at chess, like at calculation, because they are so exact and fussy about rules. Language, on the other hand, needs approximation and inference to extract "meaning" (whatever that is) together from text, context, subject matter, tone, expectations, and so on - and the computer needs yet more approximation to produce a translated version of that meaning with all the right interlocking features. Unlike chess, the game of language is played on the human home-turf of multivariate inference and approximation, so we will continue to beat the machines. But for Google's purposes, perfect translation may not even be necessary. Google succeeded in web-search partly by avoiding the exact search language of AltaVista in favor of a tool which was fast, easy to use, and displayed most of the right results in mostly the right order. Perhaps it will also be enough for Google to machine-translate most of the right words in mostly the right order, leaving to users the much harder task of extracting meaning from them. ® Bill Softky (http://www.softky.com/Bill/) has worked on dozens of science and technology projects, from the deep paradoxes of nerve cells to automatically debugging Windows source code. He hopes someday to reverse-engineer the software architecture of mammalian learning, and meanwhile works as Chief Algorithmist at an internet advertising startup. Related linksBreathless blog about Google's new translation engine (http://blog.outer-court.com/archive/2005-05-22-n83.html)
Tower of Google uses stats for translation [printer-friendly] | The Register
www.theregister.com/2007/03/29/google_translator/p... Tower of Google uses stats for translationBy John Oates Published Thursday 29th March 2007 12:44 GMT Google dreams of a world where hundreds of languages can be simultaneously translated by machines which compare texts using statistics rather than applying grammatical rules. Statistical machine translation uses a computer to compare two documents - one in the original language and one translated by a human. It finds patterns and links between the two and uses them to create its own future translations. Google has used documents from the European Commission and United Nations to feed its machines. Franz Och, who runs Google's translation team, told Reuters that early efforts impressed people with experience of machine-run translation systems. Och said: "The more we feed into the system the better it gets." Google already offers statistical translation of Arabic, Chinese and Russian. Other language translations are provided by third parties. Och repeated the Google mantra that the focus was on improving the software and that once it was working well they would look at making money from it. Dr Miles Osborne, a lecturer at the University of Edinburgh who spent last year on sabbatical at Mountain View working on the system, told the Reg: "This is quite a recent move by Google - they hired Franz Och one of leading lights in statistical translation. What you see with this system is what an academic would make if they had lots of money, support, and access to lots of machines. They have one of the world's best translators, especially for Arabic and Mandarin." Osborne said the development was important for Google because documents on the web are increasingly in languages other than English. To continue to improve its core search engine, Google would need translation software. Asked why Arabic and Mandarin were the first languages chosen, Osborne said it was down to US paranoia and homeland security. He said the US military put cash into research for translating languages from areas they considered a threat. Osborne said the US Army preferred computer-based systems because they distrusted human translators. Reuters' story is here (http://www.reuters.com/article/technologyNews/idUSN1921881520070328?feedType=RSS%20target=). Google's language tools are here (http://www.google.com/language_tools). ® Related storiesUS warns on spooky Canadian coins (11 January 2007) Welcome to the AI Nuggets Notebook!
I'm using this notebook to collect snippets of Web pages and notes related to Artificial Intelligence. The first ten or so sections correspond closely to the structure of the CSC/CPE 480 class, followed by topics not covered in detail. Towards the end you'll find applications of AI, and potential project ideas. Overview
Searching for Intelligence in Our Genes: Scientific American
www.sciam.com/article.cfm?id=searching-for-intelli... Scientific American Magazine - October 8, 2008 Searching for Intelligence in Our GenesIQ is easy to measure and reflects something real. But scientists hunting among our genes for the factors that shape intelligence are discovering they are more elusive than expectedBy Carl Zimmer In Robert Plomin’s line of work, patience is essential. Plomin, a behavioral geneticist at the Institute of Psychiatry in London, wants to understand the nature of intelligence. As part of his research, he has been watching thousands of children grow up. Plomin asks the children questions such as “What do water and milk have in common?” and “In what direction does the sun set?” At first he and his colleagues quizzed the children in person or over the telephone. Today many of those children are in their early teens, and they take their tests on the Internet. In one sense, the research has been a rousing success. The children who take the tests are all twins, and throughout the study identical twins have tended to get scores closer to each other than those of nonidentical twins, who in turn have closer scores than unrelated children. These results—along with similar ones from other studies—make clear to the scientists that genes have an important influence on how children score on intelligence tests. But Plomin wants to know more. He wants to find the specific genes that are doing the influencing. And now he has a tool for pinpointing genes that he could not have even dreamed of when he began quizzing children. Plomin and his colleagues have been scanning the genes of his subjects with a device called a microarray, a small chip that can recognize half a million distinctive snippets of DNA. The combination of this powerful tool with a huge number of children to study meant that he could detect genes that had only a tiny effect on the variation in scores. Still, when Plomin and his co-workers unveiled the results of their microarray study—the biggest dragnet for intelligence-linked genes ever undertaken—they were underwhelming. The researchers found only six genetic markers that showed any sign of having an influence on the test scores. When they ran stringent statistical tests to see if the results were flukes, only one gene passed. It accounted for 0.4 percent of the variation in the scores. And to cap it all off, no one knows what the gene does in the body. “It’s a real drag in some ways,” Plomin says. Plomin’s experience is a typical one for scientists who study intelligence. Along with using microarrays, they are employing brain scans and other sophisticated technologies to document some of the intricate dance steps that genes and environment take together in the development of intelligence. They are beginning to see how differences in intelligence are reflected in the structure and function of the brain. Some scientists have even begun to build a new vision of intelligence as a reflection of the ways in which information flows through the brain. But for all these advances, intelligence remains a profound mystery. “It’s amazing the extent to which we know very little,” says Wendy Johnson, a psychologist at the University of Minnesota. Hidden in Plain Sight Yet that kind of gut instinct does not translate easily into a scientific definition. In 1996 the American Psychological Association issued a report on intelligence, which stated only that “individuals differ from one another in their ability to understand complex ideas, to adapt effectively to the environment, to learn from experience, to engage in various forms of reasoning, to overcome obstacles by taking thought.” To measure these differences, psychologists in the early 1900s invented tests of various kinds of thought, such as math, spatial reasoning and verbal skills. To compare scores on one type of test to those on another, some psychologists developed standard scales of intelligence. The most familiar of them is the intelligence quotient, which is produced by setting the average score at 100. IQ scores are not arbitrary numbers, however. Psychologists can use them to make strong predictions about other features of people’s lives. It is possible to make reasonably good predictions, based on IQ scores in childhood, about how well people will fare in school and in the workplace. People with high IQs even tend to live longer than average. “If you have an IQ score, does that tell you everything about a person’s cognitive strengths and weaknesses? No,” says Richard J. Haier of the University of California, Irvine. But even a simple number has the potential to say a lot about a person. “When you go see your doctor, what’s the first thing that happens? Somebody takes your blood pressure and temperature. So you get two numbers. No one would say blood pressure and temperature summarize everything about your health, but they are key numbers.” Then what underlies an intelligence score? “It’s certainly tapping something,” says Philip Shaw, a psychiatrist at the National Institute of Mental Health (NIMH). The most influential theory of what the score reflects is more than a century old. In 1904 psychologist Charles Spearman observed that people who did well on one kind of test tended to do well on others. The link from one score to another was not very tight, but Spearman saw enough of a connection to declare that it was the result of something he called a g factor, short for general intelligence factor. How general intelligence arose from the brain, Spearman could not say. In recent decades, scientists have searched for an answer by finding patterns in the test scores of large groups of people. Roughly speaking, there are two possible sources for these variations. Environmental influences—anything from the way children are raised by their parents to the diseases they may suffer as they develop—are one source. Genes are another. Genes may shape the brain in ways that make individuals better or worse at answering questions on intelligence tests. Starting in the 1960s, scientists have gotten clues about the roles of genes and environment by studying twins. To see why twins are so important to intelligence researchers, imagine that a pair of identical twins are separated as babies and adopted by different parents. They have the same genes but experience different environments. If their genes have no influence at all on their intelligence test scores, then you would expect that the scores would be no more similar to each other than those of two unrelated people. Yet if genes do play a critical part in intelligence, identical twins should be more similar. “Two people with the same genes correlate as much as a person does with himself a year later,” Plomin says. “Identical twins reared apart are almost as similar as identical twins reared together.” But these similarities also take time to emerge. “By the age of 16 these adopted-away children resemble their biological parents’ IQ just as much as kids do who are reared by their biological parents,” Plomin adds. Results such as these persuaded Plomin that genes have a crucial role in intelligence, although they clearly do not act alone. “That’s what led me to say, ‘What we need to do is begin to find some genes,’” he says. Uncharted Territory Plomin could not simply scan the human genome, because it had not been mapped yet. But geneticists had identified a number of genes that, if mutated in certain ways, were associated with mental retardation. Other variations in those genes, Plomin reasoned, might produce subtler differences in intelligence. He and his colleagues compared children who scored well or poorly on intelligence tests. They looked for variants of the 100 genes that showed up unusually often in one group or the other. “We didn’t really find any there,” he says. So Plomin expanded the search. Rather than looking at a predefined set of genes, he mapped thousands of genetic markers sprinkled across the chromosomes of his subjects. If a marker turned up frequently in high- or low-scoring students, there might be an intelligence-linked gene not far away. He and his colleagues added more children to their study so that they could detect genes with weaker effects. At one point in the research, Plomin thought he had found an authentic link between intelligence and a gene known as IGF2R that encodes a growth factor receptor which is active in the brain. But when he and others tried to replicate the result, they failed. “It doesn’t look like that has panned out,” he says. Plomin suspected that he needed more genetic markers to find intelligence genes. When eggs and sperm develop, their chromosomes swap segments of DNA. The closer two segments of DNA are to each other, the more likely they are to be passed down together. But in Plomin’s early studies, millions of DNA nucleotides separated each pair of markers. It was possible that intelligence genes were so far from a genetic marker that they were sometimes getting passed down together and sometimes not. He needed a much denser set of genetic markers to reduce the chance of this happening. It was with great delight that Plomin got his hands on microarrays that could detect 500,000 genetic markers—hundreds of times more than he had previously used. He and his colleagues got cheek swabs from 7,000 children, isolated their DNA, and ran it through the microarrays. And once more the results were disappointing. “I’m not willing to say that we have found genes for intelligence,” Plomin declares, “because there have been so many false positives. They’re such small effects that you’re going to have to replicate them in many studies to feel very confident about them.” Failing to find genes for intelligence has, in itself, been very instructive for Plomin. Twin studies continue to persuade him that the genes exist. “There is ultimately DNA variation responsible for it,” he says. But each of the variations detected so far only makes a tiny contribution to differences in intelligence. “I think nobody thought that the biggest effects would account for less than 1 percent,” Plomin points out. That means that there must be hundreds—perhaps thousands—of genes that together produce the full range of gene-based variation in intelligence. Plomin doubts that some genes are specialized just for verbal skills and that other genes are just for spatial understanding. In twin studies, individuals tend to have similar scores on tests for all of those different kinds of intelligence. If genes belonged to specialized sets, a person could inherit one kind of aptitude and not the others. Plomin also surmises that his results offer some hints as to how genes influence intelligence in the brain. “If there are many genes of small effect, it’s highly unlikely that they are all going to focus in one area of the brain,” he argues. Instead the genes may be influencing a large network of brain regions. And each of those intelligence-associated genes may produce many different effects in different parts of the brain. The ultimate test of Plomin’s hypothesis will have to wait until scientists finally put together a list of genes that have an indisputable effect on how the brain works and that are associated with intelligence scores. That list may take a long time to come together, but Plomin is encouraged by new results from an entirely different line of research: a burst of new neuroimaging studies that attempt to find the mark of intelligence in the brain itself. The Shape of Intelligence In all children the cortex gets thicker as new neurons grow and produce new branches. Then the cortex thins out as branches are pruned. But in some parts of the cortex, Shaw found, development took a different course in children with different levels of intelligence. “The superclever kids started off very thin,” Shaw says. “They got really relatively thicker, but in adolescence they got thinner again very quickly.” A similar pattern has emerged from studies on adult brains. Researchers have found that people with high intelligence scores tend to have certain regions of the cortex that are larger than average. Shaw expects that some of those patterns will turn out to be the result of the environment. But these regions of the cortex tend to be the same size in twins, indicating that genes are responsible for some of the difference as well. In recent years, scientists have also published a number of studies in which they claim to have found distinctive patterns of brain functioning in people who score high on intelligence tests. Recently Haier and Rex Eugene Jung of the University of New Mexico surveyed 37 studies examining regional brain size or activity to look for an overall pattern to their results. As Plomin would have predicted, Haier and Jung found no one “intelligence spot” in the brain. Instead they identified a number of significant regions scattered around the cortex. Other studies have implicated each of these regions in different kinds of cognition. “It looks like intelligence is built on these fundamental cognitive processes, like attention and memory, and maybe language ability,” Haier says. Along with describing the gray matter tissue that makes up the cortex, these studies also find the signature of intelligence in the white matter that links distant parts of the cortex to one another. People with high intelligence tend to have tracts of white matter that are more organized than other people. “The white matter is like the wiring,” Haier says. “If you think about it, you know, intelligence really requires processing power and speed; the white matter would give it the speed; the gray matter would give it the processing power.” Haier suggests that these parts of the “intelligence network” may work differently in different people. “You can think about being very intelligent because you have a lot of speed and a lot of processing—you have both,” he says. “Or you can think about a lot of one and less of the other. All these combinations may produce the same ultimate result, so you may have two equally intelligent people, but their brains are fundamentally arriving at that behavior, however you’re measuring it, in different ways.” Haier acknowledges that these ideas are little more than speculation. Still, he argues that neuroimaging has already given scientists a far more solid understanding of intelligence. “I can predict full-scale IQ with the amount of gray matter in a small number of areas,” he says. Haier suspects that in the near future, 10 minutes in a magnetic resonance imaging scanner may reveal as much about high school students as four hours taking an SAT exam. Some psychologists are not quite ready to take that step. They do not think IQ and g should be endowed with a deeper significance than they deserve. For one thing, there is much beyond the life of the mind than mentally rotating cubes and completing analogies. “I think human intelligence is multifaceted and very complex,” the University of Virginia’s Turkheimer says. Unfortunately, he adds, barely any work has been done on other facets of intelligence. “We can use g for a lot of useful things, but I don’t believe it follows from that that human intelligence is a unitary thing called g that we can find in a literal way in the brain,” he says. Longitude and latitude are useful, too, for navigation, he notes, but that does not mean there is actually a grid carved into the earth. Johnson of the University of Minnesota defends the g factor as tapping into something important in the brain, but, like Haier, she does not think it is a one-size-fits-all general intelligence. “While there is something general about intelligence, what makes my intelligence general is not the same thing that makes your intelligence general or any other person’s,” she says. “Our brains are plastic enough that we put together, each of us, a different kind of general intelligence.” Pinpointing the role genes have in producing different kinds of intelligence will no doubt be very difficult. And it is entirely possible that a list of intelligence-linked genes may include many that do not actually have brain-specific functions. Turkheimer offers the following thought experiment: Imagine a gene that is related to the width of a woman’s birth canal. Women who carry a gene for a narrow birth canal tend to have more trouble in labor, and their babies run a higher risk of being oxygen-deprived. As a result, their babies, on average, have IQ scores a couple points below those who have a different version of the gene. And some of those children will also carry the narrow-canal gene. “These babies are going to have a gene that’s correlated with low IQ,” Turkheimer says. “So do you conclude that that’s an IQ gene? Well, not really; it’s a birth-canal gene. The ways that genes could correlate with IQ are so variable that it’s almost impossible to know.” Turkheimer’s own research illustrates another kind of complexity in the link from genes to intelligence: genes do not act in isolation from the environment. In fact, the same gene can have different effects in different environments. Turkheimer was led to this realization when he noticed that the large twin studies on intelligence contained few poor children. “Very poor people don’t have the time or the resources or the interest to do volunteer studies,” he says. Other databases, Turkheimer discovered, have more poor children in their ranks. He was able to analyze the test scores of hundreds of twins, taking into consideration their socioeconomic status—a category based on factors that include a family’s income and the education level of the parents. He found that the strength of genes’ effect depended on the socioeconomic status of the children. In children from affluent families, about 60 percent of the variance in IQ scores could be accounted for by genes. For children from impoverished families, on the other hand, genes accounted for almost none. Turkheimer and his colleagues published these results in 2003; in May 2007 they replicated the pattern with another database. Instead of comparing IQ scores, the researchers looked at how 839 pairs of twins fared on the National Merit Scholarship Qualifying Test in 1962. Once more genes played little role in the variance of scores among poor children and played a far stronger one in more affluent children. Turkheimer posits that poverty brings with it powerful environmental forces that may shape intelligence from the womb through school and onward. But when children grow up in the relative stability of an affluent home, gene-based differences can begin to make themselves felt. As if this complexity was not enough, scientists are also finding that genes, in turn, can alter the effect that the environment has on our intelligence. Last year British scientists found an association between breast-feeding and a boost in IQ test scores—but only if children carried a particular variant of a particular gene. If they carried another variant, breast-fed children scored no differently than children who drank formula. Genes may also influence behavior in ways that influence how intelligence develops. “People create their own environments,” Johnson says. “If you see a kid who’s really interested in art or math, you’re just more likely to go out and get him a math book or some crayons. So they’ll practice it and become more different from the kid who doesn’t have the math book. Parents respond to what the kid does. Our models don’t measure that well at all.” This effect might explain one of the most puzzling patterns in twin studies on intelligence: how the influence of genes becomes stronger on test scores as people get older. Genes may affect how people mold their intellectual environment. Choosing to seek out new experiences, reading books and engaging in conversations may alter the brain. And as children grow up and take over control of their own lives, this effect may get stronger. “Intelligence is kind of an emergent property of the brain,” Shaw says. “The idea that you’re born with 15 genes, and they set in stone how intelligent you’re going to be and how your brain is going to develop, is almost certainly wrong.” Why Study Intelligence? “It’s very important as we enter the 21st century to maximize and optimize education for people,” he argues. “That’s what’s at stake.” By understanding people’s genetic profiles, Plomin suggests, it may be possible to find the best ways to foster learning. If, as he anticipates, microarray studies finally do reveal intelligence genes, it may then be possible to test children for which versions they carry. And for some psychologists, it is enough that intelligence is such an intriguing part of human nature. “Intelligence and intelligence test scores are in many ways the best predictor in all of psychology,” Turkheimer says. “That’s what makes it fascinating. If you know my SAT scores and you want to know how I’m going to perform at practically anything, those SAT scores are far from a perfect predictor, but they’re far better than knowing about my personality. Intelligence really works. There’s this palpable psychological quality that allows you to make predictions about humans but gets very slippery when you try to tie it down in concrete numerical ways. So it’s just a very interesting scientific problem.” For charts and graphs related to this story , click the images below to enlarge
Note: This article was originally titled, "The Search for Intelligence". Computer Science:Artificial Intelligence - Wikibooks, collection of open-content textbooks
en.wikibooks.org/wiki/Computer_Science:Artificial_... Computer Science:Artificial IntelligenceFrom Wikibooks, the open-content textbooks collectionJump to: navigation, search
[edit] IntroductionWhile no consensual definition of Artificial Intelligence (AI) exists, AI is broadly characterized as the study of computations that allow for perception, reason and action. We can divide the cognitive world into 3 domains - Intelligent, Normal and Below Normal. The average human being does Normal things. Intelligence is always demonstrated in NEW domains. The properties of a New domain are that the patterns in these domains are ambiguous and are incomplete. Therefore, Intelligence can be better understood as the ability to work efficiently with Incomplete, Ambiguous Patterns. Artificial Intelligence extends the computing power from an algorithmic domain to this Incomplete, Ambiguous patterns domain. (Algorithms have the property of Definiteness which cannot be satisfied in this domain.) There are two major views of work in the field of AI: some people view AI as the quest for mechanisms, or algorithms, in which their output would be considered intelligent if performed by a human (i.e., a chess playing computer). Others view AI as a scientific discipline to understand the human mind, by modelling the cognitive processes humans go through. With a bit of oversimplification, AI is:
Main techniques include:
AI has influenced a number computer scientific areas:
[edit] SearchSearch is a classical AI topic. Many AI problems reduce to searching for solutions in the problem state space. The basic idea:
Example: symbolic integration To find this integral: Example: 8-puzzle If a bunch of tiles are arranged like this:
And this is the correct order:
and the [edit] Heuristic searchYou can use DFS (Depth-first search) or BFS (Breadth-first search) to explore a graph, but it may require O(bm) node visits, but if you want to search more efficiently, it might be better to check out the promising leads first. With a heuristic search, at each step of the way, you use an evaluation function to choose which subtree to explore. You can think of a heuristic search like a "best-first" search. It is important to remember that a heuristic search is best at finding a solution that is "good enough" rather than the perfect solution. Some problems are so big that finding the perfect solution is too hard. To do a heuristic search you need an evaluation function that lets you rank your options. [edit] Sliding tile puzzle aka 8-puzzle ExampleThis website has an example of a sliding tile puzzle if you are unfamiliar with what these are. Sliding tile puzzles (or 8-puzzles) provide a problem that can be solved more efficiently with a heuristic search (like A*) vs. DFS and BFS. For the 8-puzzle, we could use this as a simple evaluation function: f = g + w where g is the distance from the root, and w is the number of misplaced tiles, not counting the The evaluation function f = g + w would score the initial state
as 4 = 0 + 4 A better heuristic is known as the Manhattan distance. This heuristic value is determined by the sum of the distance for each tile from its goal location. This heuristic is admissible, such that it does not overestimate the cost to reach the goal node. TODO: draw search tree for 8-puzzle and show heuristic search path solution, vs. DFS and BFS search solution. The above evaluation function is just one possible guess. We could play with other evaluation functions, like: f = g + p, where g is the same, and p is the sum of distances that each tile is from the root. Or we could use f = g + p + 3S, where g and p are the same, and S is the sum of sequence scores Σs. s is calculated for each tile like so:
[edit] Tic Tac Toe and minimaxto be done later n,mn,mn,mn [edit] Knowledge discovery and machine learning[edit] Probability is probably going to be interestingExample 1: tossing a coin once ω1 = HEADS ω2 = TAILS
X(ω) can be a function like the number of heads. X(ω1) = 1, and X(ω2) = 0. Example 2: tossing a coin twice.
Potiential outputs from our X(ω) function:
On to probability. P(HH) should return the probability of getting two heads. Often we denote P(X(ω) = x) as P(x). [edit] Logarithms refresherThe logarithm L of a number N, in a given base B, is such that N = BL For instance, make N = 8 and B = 2. Then, you would have L = 3, as you can see: log28 = 3 Another example:
[edit] Emacs SLIME appendixThe recommended way to use lisp (according to the #lisp crew) is to use SBCL and emacs with the SLIME extension. Below are some useful commands: Command what it does M-x slime starts slime session C-x b switches to another buffer C-c C-l load current file into SLIME C-c C-k compile current file C-c C-c compile current function C-x C-s save current buffer C-x C-c quit emacs C-c C-z jump to slime buffer C-x C-f open a new buffer C-x 2 view two frames, horizontally stacked. C-x 1 return to a one-frame view C-x o move to other frame [edit] Common lisp resources[edit] useful links
[edit] a few lisp coding examplesCL-USER> (loop for e in '(a b c)
do (format t "e is ~a.~%" e)
collect e)
e is A.
e is B.
e is C.
(A B C)
CL-USER> (let ((x 0)
(y 1)
(z 2))
(format t "x: ~a y: ~a z: ~a~%" x y z)
(+ x y z))
x: 0 y: 1 z: 2
3
CL-USER> (if (< 3 1) 0 1)
1
CL-USER> (if (< 3 5) 1 2)
1
CL-USER> (cond ((< 3 1) 1)
((< 9 10) 2)
(t 3))
2
[edit] class notes on lispLisp is made of symbolic expressions. Everything descends from them.
A lisp program is an ordered set of lists. In lisp, programs and data have the same form. Programs can be manipulated by programs. (+ 8 3) In the above computation, + is an operator, and 8 and 3 are arguments. Some example computations: CL-USER> (- 8 3) 5 CL-USER> (- 8 3 4) 1 CL-USER> (1+ 8) 9 CL-USER> (1- 8) 7 CL-USER> (* 8 (+ 3 7)) 80 CL-USER> (expt 2 3) 8 CL-USER> (expt 3.3 2.2) 13.827086 CL-USER> (sqrt 9) 3 CL-USER> (abs -5) 5 [edit] Symbols and setqx is a symbol, which is sort of like a variable. (setq x 5) ; assigns 5 to symbol x. setq does not evaluate its first argument. Symbols and variables are different; e.g., x is a symbol throughout the program. However, x can be bound to different variables, e.g., a global and a local. The global x may have a value of 5, while the local x has the value 8. setq can make multiple assignments: CL-USER> (setq x 5 y 6 y 7) 7 [edit] quoteCL-USER> (quote (a b c)) (A B C) CL-USER> '(a b c) (A B C) CL-USER> (setq x '(+ 3 4)) (+ 3 4) CL-USER> (setq x (+ 3 4)) 7 CL-USER> (setq x '(+ 3 (* 4 5))) (+ 3 (* 4 5)) CL-USER> (setq x `(+ 3 ,(* 4 5) (- 6 7))) (+ 3 20 (- 6 7)) [edit] setq vs setCL-USER> (setq x 'y) Y CL-USER> x Y CL-USER> (set x 'z) Z CL-USER> x Y CL-USER> y Z set, rather than setq, evaluates the first argument. setq stands for set quote. [edit] setf, remove, consassignment of lists: CL-USER> (setf friends '(dick jane)) (DICK JANE) CL-USER> friends (DICK JANE) CL-USER> (setf enemies '(grinch ghost)) (GRINCH GHOST) CL-USER> (setf friends (remove 'ghost enemies)) (GRINCH) CL-USER> (setf friends (cons 'ghost friends)) (GHOST GRINCH) CL-USER> friends (GHOST GRINCH) first, second, third, etc. CL-USER> (first '(a b c)) A CL-USER> (second '(a b c)) B CL-USER> (third '(a b c)) C car and cdr: old-school flavor. CL-USER> (cdr '((a b) c)) (C) CL-USER> (car (cdr '(a b c))) B You can combine car and cdr: CL-USER> (car (cdr '(a b c))) B CL-USER> (cadr '(a b c)) B CL-USER> (caddr '(x y z w)) Z CL-USER> (car (cdr (cdr '(x y z w)))) Z Fun with cons: CL-USER> (cons 'a '(b c)) (A B C) CL-USER> (cons '(a b) '(c d)) ((A B) C D) CL-USER> (cons 'a (cons 'b (cons 'c NIL))) (A B C) CL-USER> (list 'a 'b 'c) (A B C) CL-USER> (append '(a b) '(c d)) (A B C D) CL-USER> cons, list, and append do not alter symbol values. CL-USER> (setq x 'a L '(b c)) (B C) CL-USER> (cons x L) (A B C) CL-USER> L (B C) CL-USER> (setq x 'a L '(b c)) (B C) CL-USER> (setq L (cons x L)) (A B C) CL-USER> L (A B C) Inserting an element at the end of a list: CL-USER> (append '(a b) '(c)) (A B C) CL-USER> (setq x 'c) C CL-USER> (append '(a b) (list x)) (A B C) CL-USER> (length '(a b c)) 3 CL-USER> (length '((a b) (c d))) 2 Reverse only reverses the order of the top list elements: CL-USER> (reverse '(a b c)) (C B A) CL-USER> (reverse '((a b) (c d))) ((C D) (A B)) [edit] alists and assocCL-USER> (setf apple '((ako fruit)
(color red)
(shape sphere)))
((AKO FRUIT) (COLOR RED) (SHAPE SPHERE))
CL-USER> (assoc 'color apple)
(COLOR RED)
[edit] three rules of evaluation1. Symbols are replaced by their current bindings CL-USER> (setq x 2) 2 CL-USER> (+ x x) 4 2. Lists are evaluated as function calls. 3. Everything else evaluates to itself. It may not look particularly significant, but the above illustrates an important characteristic. You can write symbols rather than only reserved words and variables. [edit] Generations of computer languagesMachine language -> assembly -> high-level languages -> very high-level languages -> symbolic languages [edit] FunctionsCL-USER> (defun addthree (x) (+ x 3)) ADDTHREE In addthree, x is a formal parameter, meaning that x is in the function definition, rather than the function call. CL-USER> (defun sum-average (x y)
(setq sum (+ x y))
(/ sum 2))
SUM-AVERAGE
sum is a non-formal parameter. non-formal parameters are free rather than bound. x and y are formal parameters, which are bound. free parameters are global, and bound parameters are local. CL-USER> (sum-average 3 5) 4 CL-USER> sum 8 CL-USER> (sum-average 10 20) 15 CL-USER> sum 30 Notice that sum in the top level has its value changed. Use free symbols with great care. It is conventional to set off global parameters with asterisks, like *PS*. CL-USER> (defun sum-ave-caller (sum x y)
(sum-average x y) sum)
SUM-AVE-CALLER
CL-USER> (sum-ave-caller 0 10 20)
0
CL-USER> sum
30
Take a look at the above code carefully. It may be confusing but the code listed below is equivalent. CL-USER> (defun sum-ave-caller (sum_name_substitute x y)
(sum-average x y) sum_name_substitute)
SUM-AVE-CALLER
CL-USER> (sum-ave-caller 0 10 20)
0
CL-USER> sum
30
CL-USER> sum_name_substitute
ERROR
[edit] TheoryThe discipline of artificial intelligence potentially provides answers to unsolved problems in neuroscience because AI must devise and test theories of how mind emerges from brains. I hope you got the idea. Artificial Intelligence - Wikibooks, collection of open-content textbooks
en.wikibooks.org/wiki/Artificial_Intelligence Artificial IntelligenceFrom Wikibooks, the open-content textbooks collectionJump to: navigation, search
Welcome to the Wikibook about Artificial Intelligence. Before starting on the book, a basic layout and method should be decided on. See the discussion page for this.
[edit] IndexThe following is a first proposal for a basic layout. This is not yet complete, ideas are welcome. Discuss on the talk page or just add them here. The book is laid out into 4 sections, with increasing detail and complexity. Each section contains a number of chapters. In addition to regular chapters, there are case-study chapters, that investigate full and complex AI systems using several techniques from the regular chapters (as well as perhaps some new ones) [edit] Introduction
Some application of AI: AI has a wide spectrum of applications including natural language processing, search engines, medical diagnosis, bioinformatics and cheminformatics, detecting credit card fraud, stock market analysis, classifying DNA sequences, speech and handwriting recognition, object recognition in computer vision, game playing, machine learning and robot locomotion. Some scientists and futurists predict that in near future with AI we can make digital human, artificial life and artificial immortality.
[edit] Section 1[edit] General concepts
[edit] Section 2Basic AI topics
[edit] Section 3More advanced topics and techniques in AI
Section 4: Highly specific AI topics and techniques.
Section 5 A.I Circuits and algorithms
Albtraum ComputerIst das menschliche Gehirn nur eine Maschine aus Fleisch? Ein entscheidender Artikel des selbsternannten "Ketzers der Informatik"aus der ZEIT vom Januar 1972Von Joseph Weizenbaum Unsere Zivilisation steht heute am Anfang einer schweren geistigen Krise. Akademiker, Industrielle und Journalisten beschäftigt die Möglichkeit, daß der Computer irgendwie beweisen wird, „das Gehirn sei lediglich eine Maschine aus Fleisch". Allein eine solche These zu erwägen bedeutet, den Nutzen der Freiheit des Menschen, seiner Würde und seiner Autonomie in Frage zu stellen. Wie hat der Computer zur Entstehung dieser betrüblichen Sachlage beigetragen? Wir müssen uns darüber klar sein, daß ein Computer nichts ist ohne ein Programm. Ein Programm ist im Grunde die Umwandlung eines Computers in einen anderen, welcher autonom ist und — in einem sehr realen Sinne — ein Verhalten besitzt. Programmiersprachen beschreiben dynamische Prozesse. Wir können Programmodelle für jeden beliebigen Aspekt der realen Welt, der uns interessiert, konstruieren, und wir können diese Modelle arbeiten lassen. Doch müssen wir uns immer vor Augen halten, daß ein Computermodell eine Beschreibung ist, die da arbeitet. Wenn wir gemeinhin davon sprechen, daß ein Modell von B sei, so meinen wir damit, daß eine Theorie über irgendwelche Aspekte des Verhaltens von B zugleich eine Theorie der gleichen Aspekte des Verhaltens von A sei. Eröffnen uns die Programmiersprachen neue Möglichkeiten des Ausdrucks, so entbinden sie uns doch keineswegs von der Verpflichtung, verfechtbare Theorien aufzubauen. Denn auch Fehler können mit äußerster Präzision und Beredsamkeit behauptet werden, aber dadurch werden sie noch nicht in Wahrheit verwandelt.
Die fehlende Unterscheidung zwischen Beschreibungen (selbst solchen, die „funktionieren") und Theorien erklärt weitgehend, warum diejenigen, die den Menschen nicht als Maschine sehen wollen, in die Defensive gedrängt worden sind. Neuere Fortschritte im Verstehen natürlicher Sprachen durch Computer sind ein gutes Beispiel. Die Linguisten Halle und Chomsky am Massachusetts Institut für Technologie haben lange an einer Sprachtheorie gearbeitet, der jedes beliebige Modell sprachlichen Verhaltens genügen muß. Während ihre Erfolge von ihresgleichen gefeiert und bisweilen von anderen verdammt werden, sind sie weiterhin von zwei Dingen überzeugt: Viel wichtiger ist aber, daß sie ihr Arbeitsgebiet mit großer Ehrfurcht und tiefer persönlicher Demut betrachten. Informatiker, die sich mit dem Verstehen natürlicher Sprachen durch Computer beschäftigen, mögen ebenso bescheiden sein. Aber sie arbeiten in einer gewöhnlich als Performanz-Modus bezeichneten Arbeitsweise. Darin zählt nicht die Ausarbeitung einer Theorie, sondern die Leistung von Systemen. Verführung und Irreführung Ich habe keinerlei Zweifel, daß am Ende dieses Jahrzehnts Computersysteme existieren werden, mit denen Spezialisten wie etwa Ärzte, Chemiker und Mathematiker in natürlicher Sprache verkehren werden. Und sicherlich wird ein Teil dieser Errungenschaften auf anderen Erfolgen aufbauen, wie zum Beispiel auf der Simulation von Erkenntnisprozessen durch Computer. Daher ist es verständlich, wenn sich gewisse Täuschungen anzusiedeln und auszubreiten beginnen. Aber allein die Frage „Hat der Computer das Wesen des Menschen erfaßt?" ist eine Irreführung und damit eine Falle. Denn die eigentliche Frage „Versteht der Mensch das Wesen des Menschen?" hat nichts mit Technologie zu tun und ganz sicher auch nichts mit irgendeinem technischen Gerät. Wir haben technologische Metaphern - „Mythen der Maschine" - und die Technik selbst so tief in unsere Gedankenprozesse eindringen lassen, daß wir schließlich an die Technologie sogar die Aufgabe, Fragen zu formulieren, abgegeben haben. Kluge Menschen empfinden zu Recht, daß große Datenbänke und riesige Computernetze den Menschen bedrohen. Aber sie überlassen es der Technologie, die entsprechende Frage zu formulieren. Wo ein einfacher Mann fragen würde: „Brauchen wir diese Dinge?", fragt die Technologie: „Welche elektronische Zauberei macht sie ungefährlich?" Wo ein einfacher Mann fragt: „Ist das gut?", fragt die Technologie: „Wird das funktionieren?" Auf diese Weise wird die Wissenschaft und sogar die Weisheit zu dem, was Technologie und vor allem der Computer handhaben können. Die Frage „Ist das Gehirn lediglich eine Maschine aus Fleisch?" ist typisch für die Art von Fragen, die aus einer technologischen Mentalität formuliert und tatsächlich nur in ihr formulierbar sind. Sobald sie als rechtmäßig zugelassen ist, beginnen Streitgespräche, was ein Computer „im Prinzip" kann oder nicht kann, und diese Streitgespräche werden dann selbst rechtmäßig. Aber die Zulässigkeit der technologischen Frage braucht nicht von vornherein anerkannt zu werden. Statt dessen kann eine menschliche Frage gestellt werden. Systeme ohne Autoren Der Erfolg der Technik und einiger technologischer Erklärungen hat uns dazu verleitet, der Technologie die Formulierung wichtiger Fragen an unserer Statt zu erlauben — Fragen, deren Form schon die Anzahl der Freiheitsgrade in unserem Entscheidungsraum ernsthaft einschränkt. Wer immer die Frage stellt, bestimmt in starkem Maße die Antworten. In diesem Sinne ist die Technologie und speziell die Computertechnologie ein sich selbst erfüllender Alptraum geworden, der an die Frau erinnert, die davon träumt, vergewaltigt zu werden, und ihren Angreifer bittet, nett zu ihr zu sein. Er aber antwortet: „Es ist Ihr Traum, gute Frau." Wir müssen einsehen, daß die Technologie unser Traum ist und daß wir es sind, die schließlich entscheiden, wie er enden wird. Die Computerrevolution muß die Würde und Autonomie des Menschen weder in Frage stellen noch braucht sie es, sondern sie ist eine Art pathologisches Phänomen, das den Menschen dazu bewegt, ihm unberechtigte, höchst schädliche Interpretationen abzuringen. Sobald wir uns klarmachen, daß unsere Visionen, möglicherweise Alpträume, die Wirkung unserer eigenen Werke auf uns und unsere Gesellschaft bestimmen, so wird ihre Bedrohung sicherlich vermindert. Das bedeutet aber nicht, daß dieses Bewußtwerden bereits alle Gefahren ausschließt. Zum Beispiel gibt es, außer der aushöhlenden Wirkung einer technologischen Mentalität auf das menschliche Selbstverständnis, unmittelbare Angriffe auf die Freiheit und Würde des Menschen, in denen die Computertechnologie eine kritische Rolle spielt. Wir haben bereits eine Maschine (Dendral), die über mehr chemisches Wissen verfügt als viele Doktoren der Chemie, und eine andere (Mathlab), die mehr Wissen über Angewandte Mathematik besitzt als die meisten Mathematiker dieses Zweiges. Die Kenntnisse beider können vor dem Hintergrund der Theorien, von denen sie abgeleitet wurden, bewertet werden. Wenn der Anwender von Mathlab das aus einer bestimmten Funktion berechnete Integral für falsch hält, so kann er, einmal abgesehen von möglichen Programmfehlern, nicht in Übereinstimmung mit der mathematischen Theorie der Integration sein. Er streitet dann nicht mit der Maschine oder dem Programmierer, sondern mit einer bestimmten mathematischen Theorie. Was aber ist mit den vielen Programmen, auf die sich das Management in Regierung und Militär verläßt, von denen man keineswegs behaupten kann, daß sie auf erklärbaren Theorien beruhen, die vielmehr statt dessen ein riesiges Flickwerk von Programmiertechniken sind, aneinandergeknüpft, damit sie funktionieren? In unserem Eifer, jeden technischen Fortschritt auszunutzen, fügen wir eiligst das, was wir bei der maschinellen Manipulation des Wissens solcher auf bestimmten Theorien basierenden Systeme gelernt haben, in dieses Flickwerk ein. Es „funktioniert" dann besser. Nun wird es ungemein wichtig zu verstehen, wie solche Systeme wirklich konstruiert sind. Ich denke da an Systeme zur Auswahl von Kriegszielen, wie sie in Vietnam benutzt werden, an die Kriegsspiele im Pentagon und so weiter. Diese oftmals gigantischen Systeme werden von Programmierteams in Zeitspannen häufig von vielen Jahren zusammengesetzt. Wenn dann das System schließlich wirklich in Gebrauch genommen wird, sind die meisten früheren Programmierer nicht mehr da oder haben sich anderen Aufgaben zugewandt. Und genau dann, wenn solche riesigen Systeme endlich benutzt werden, kann weder eine einzelne Person noch ein kleines Team von Spezialisten ihre inneren Arbeitsabläufe überblicken. Dies hat Folgen: Die Schwelle der Komplexität, jenseits derer diese Erscheinung auftritt, ist schon von vielen existierenden Systemen einschließlich einiger Kompilations- und Operationssysteme überschritten worden. So mag zum Beispiel niemand gewisse Operationssysteme für manche großen Maschinen, aber sie können weder wesentlich geändert noch können sie abgeschafft werden. Zu viele Menschen sind von ihnen abhängig. Ein plumpes Betriebssystem ist unbequem. Das an sich ist nicht schlimm. Auf einem anderen Blatt aber steht, daß man sich immer mehr auf Supersysteme verläßt, die vielleicht als Hilfe bei Analysen und Entscheidungen gedacht waren, seitdem aber das Wissen ihrer Benutzer überfordern und doch für sie unentbehrlich geworden sind. Im modernen Krieg ist es dem Soldaten, etwa dem Bomberpiloten, vertraut, in einer riesigen psychischen Entfernung von seinen Opfern zu operieren. Er ist nicht verantwortlich für verbrannte Kinder, denn er sieht ja niemals ihre Dörfer, seine Bomben und gewiß nicht die brennenden Kinder. Die moderne technologische Rationalisierung von Krieg, Diplomatie, Politik und Handel wie etwa in den Computerspielen — wirkt sich sogar noch heimtückischer auf die Politik aus: Die Politiker haben nicht nur ihre Verantwortung zur Entscheidung einer ihnen unverständlichen Technologie übertragen, wobei sie die Illusion aufrechterhalten, daß sie, die Politiker, die politischen Fragen stellen und beantworten; vielmehr ist die Verantwortung selbst verdunstet. Kein menschliches Wesen ist mehr verantwortlich für das, „was die Maschine sagt". Daher kann es weder richtig noch falsch geben, keine Rechtsfrage, keine Theorie, mit der man übereinstimmen oder nicht übereinstimmen kann, und schließlich keine Grundlage, auf der man in Frage stellen könnte, „was die Maschine sagt". Mein Vater pflegte sich auf die letzte Autorität zu beziehen, indem er zu mir sagte: „Es steht geschrieben." Aber da konnte ich lesen, was geschrieben steht, konnte mir einen Autor vorstellen, konnte seine Wertmaßstäbe rekonstruieren und schließlich zustimmen oder ablehnen. Dagegen haben die Systeme im Pentagon und ihre Gegenstücke überall in unserer Kultur in einem sehr realen Sinn keine Autoren. Sie erlauben daher keine Anwendung unseres Vorstellungsvermögens, die schließlich zu menschlicher Beurteilung führen könnte. Kein Wunder, daß Menschen, die Tag um Tag mit solchen Maschinen leben und von ihnen abhängen, glauben, daß Menschen lediglich Maschinen seien. Sie spiegeln damit wider, was sie selbst geworden sind. Die potentiellen, tragischen Wirkungen solcher Systeme auf die Gesellschaft sind noch schlimmer, als es auf den ersten Blick scheinen könnte. Und es sind die Nebenwirkungen, nicht die direkten Effekte, auf die es am meisten ankommt. Raffinierte Nebenwirkungen Die Idee raffinierter indirekter Nebenwirkungen einer Technologie auf die Gesellschaft läßt sich an der Erfindung des Mikroskops nachweisen. Zur Zeit seiner Entdeckung, in der Mitte des 17. Jahrhunderts, wurde Krankheit allgemein als eine Strafe verstanden, die Gott einem einzelnen auferlegte. Das Mikroskop befähigte den Menschen zur Erkennung von Mikroorganismen und schuf damit die Voraussetzungen für die Erregertheorie der Krankheiten. Daneben führte die überraschende Entdeckung extrem kleiner lebender Organismen zu der Vorstellung einer kontinuierlichen Lebenskette, die wiederum eine notwendige geistige Voraussetzung für die Entwicklung des Darwinismus war. Sowohl die Theorie der Krankheitserreger als auch die Evolutionstheorie veränderten die Vorstellungen des Menschen von seinem Vertrag mit Gott und damit sein Selbstverständnis. Politisch haben diese Ideen die Macht der Kirche verringert und bisher unangreifbare Autoritäten in Frage gestellt. Es ist sinnvoll zu fragen, ob der Computer ähnliche Veränderungen im Selbstverständnis des Menschen bewirken wird. Wie ist die psychologische Auswirkung auf die Individuen einer Gesellschaft, in der anonyme, also nicht verantwortungsfähige Kräfte die großen Tagesfragen stellen und den Bereich möglicher Antworten abgrenzen? Es kann nicht überraschen, wenn eine große Anzahl aufnahmebereiter Menschen in einer solchen Gesellschaft ihre Ohnmacht erkennen und sich in jene blinde Wut treiben lassen, die oft solche Erfahrungen begleitet. Computergestützte Wissenssysteme werden mehr oder weniger unveränderbar (abgesehen davon, daß sie wachsen können). Weil sie Abhängigkeit erzeugen und nach Überschreiten einer gewissen Schwelle nicht mehr aufgegeben werden können, besteht die große Gefahr, daß sie von einer Generation zur nächsten vererbt werden und dabei immer weiter wachsen. Sicherlich, auch der Mensch überträgt seine Erfahrungen von einer Generation zur nächsten. Weil er aber sterblich ist, ist diese Übermittlung über die Generationen zugleich ein Prozeß des Filterns und der Vervollkommnung. Der Mensch überträgt nicht bloßes Wissen, er regeneriert es kontinuierlich. Soviel wir auch das Verschwinden alter Kulturen betrauern mögen, so wissen wir doch, daß die Größe des Menschen ebensosehr in der Evolution seiner Kulturen wie in der Evolution seines Hirns begründet ist. Die unkluge Anwendung immer größerer und immer komplexerer Computersysteme könnte diesen Prozeß durchaus zum Stillstand bringen. Sie könnte die Ebbe und Flut der Kulturen durch eine Welt ohne Werte ersetzen, eine Welt, in der alles Wesentliche vor langer Zeit bestimmt und für alle Zeiten festgehalten worden ist. Der meiste Schaden, den der Computer potentiell zur Folge haben könnte, hängt weniger davon ab, was der Computer tatsächlich machen kann oder nicht kann, als vielmehr von den Eigenschaften, die das Publikum dem Computer zuschreibt. Der Nichtfachmann hat überhaupt keine andere Wahl, als dem Computer die Eigenschaften zuzuordnen, die durch die von der Presse verstärkte Propaganda der Computergemeinschaft zu ihm dringen. Daher hat der Informatiker die enorme Verantwortung, in seinen Ansprüchen bescheiden zu sein. Diesen Rat brauchte ich nicht einmal auszusprechen, wenn die Informatik eine Tradition der Gelehrsamkeit und Kritik besäße wie die der etablierten Wissenschaften. Beim gereiften Wissenschaftler ist gerade seine Demut die Quelle seiner Stärke. Ich betrachte es als eine der wichtigsten Aufgaben eines Fachbereichs für Informatik an einer Universität, diese Art von Demut den Studenten einzuflößen, insbesondere durch das Beispiel der Lehrenden. Darüber hinaus muß sich der Informatiker stets bewußt sein, daß seine Instrumente ungeheuerlich verstärkende Wirkungen haben können, sowohl direkt als auch indirekt. Ein Fehler in einem Programm kann ernsthafte Folgen haben, sicherlich auch den Verlust von Menschenleben. Ich nenne ein Beispiel: Am 11. September 1971 verursachte ein Programmfehler die gleichzeitige Zerstörung von 117 Stratosphären-Wetterballons, deren Instrumente von einem Erdsatelliten überwacht wurden. Ein ähnlicher Fehler in einem militärischen Kommando- und Kontrollsystem könnte ein Rudel nuklearbewaffneter Geschosse starten. Nur die allgemeine Zensur verhindert, daß wir erfahren, wie viele solcher Ereignisse mit nichtnuklearen Waffen bisher eingetreten sind. Der Informatiker hat daher die schwerwiegende Verantwortung, die Fehlbarkeit und Begrenztheit der Systeme, die er entwerfen kann, äußerst klarzumachen. Gerade die Wirkungsmöglichkeiten seiner Systeme sollten ihn zögern lassen, bereitwillig seinen Rat zu erteilen, und sollten ihn veranlassen, den Wirkungskreis seiner geplanten Arbeit einzuschränken. Die Grundfrage des Informatikers ist die gleiche, die sich jeder Wissenschaftler, ja, jeder Mensch stellen muß. Sie lautet nicht „Was soll ich tun", sondern vielmehr „Was soll ich sein". Ich kann diese Frage nur für mich selbst beantworten. Aber wenn die Technologie ein Alptraum mit anscheinend eigener unausweichlicher Logik ist, dann ist sie unser Alptraum. Der Mensch kann, Mut und Einsicht vorausgesetzt, der Technologie das Vorrecht absprechen, Menschheitsfragen zu stellen. Man kann menschliche Fragen stellen und darauf menschenwürdige Antworten finden. Zum ThemaDIE ZEIT 11/2008:
»Heben und Rollen« ZEIT online 11/2008:
Joe Weizenbaum, freier Geist images.zeit.de/text/online/2008/11/Joseph-Weizenbaum-Nachruf
images.zeit.de/text/online/2008/11/Joseph-Weizenba... Joe Weizenbaum, freier GeistProgrammieren heißt, die Gesellschaft zu gestalten. Also muss Informatik kritisch sein, und sie ist nichts ohne das kultivierte Gespräch. Erinnerung an einen Intellektuellen, der es pflegte.Von Gero von Randow Er liebte das deutsche Wort „Quatsch“ und benutzte es oft. Joseph Weizenbaum, geboren 1923 in Berlin, war ein weiser und köstlicher Quatschbekämpfer, der wusste, dass Lachen frei macht. Frei zu denken, sich nicht von den Moden - heute heißt es „Hypes“ - der sogenannten Informationsgesellschaft verblöden zu lassen, das propagierte er, und sein Humor nährte sich aus bitterem Ernst. Die Weizenbaums waren Juden und flohen 1936 vor den Nazis. Die Familie fand ihre neue Heimat, und Freiheit, in Amerika; gut möglich, dass diese Erfahrung als Hintergrund der freien kritischen Rede zu sehen ist, die Joseph Weizenbaum pflegte. Sie galt vor allem den Mythen des computertechnischen Fortschritts, und wer heute ernsthaft über die Informatisierung nachdenken will, sollte dies nicht tun, ohne Weizenbaums Kritik zu studieren. Joseph Weizenbaum war Mathematikstudent, als die ersten Computer entstanden, und er warf sich fort auf die Entwicklung dieser - ja, was waren das für Dinger? Logikmaschinen, so viel war klar. Rechner, das allemal. Aber auch „Elektronengehirne“, wie es zu jener Zeit hieß? Die Jagd nach „Künstlicher Intelligenz“ (KI) sollte bald beginnen, und Weizenbaum nahm von Anbeginn daran teil. Allerdings hatte er einen anderen Bildungshintergrund als seine akademischen Kollegen. Weizenbaum war nicht nur mit den formalen Grundlagen der Wissenschaft vertraut, die später „Informatik“ genannt werden sollte, sondern er kannte sich auch in der philosophischen Untersuchung von Begriffen wie „Bedeutung“ aus. Und es beginnt die Geschichte die schon oft erzählt wurde, hier in der Kurzfassung: Weizenbaum schreibt ein Programm, das eingetippte Sätze in natürlicher Sprache klassifizieren konnte, um darauf mit eigenen Sätzen zu antworten: einen sogenannten „Parser“. So etwas ist zu diesem Zeitpunkt, Mitte der sechziger Jahre, noch avantgardistisch; wenige Jahre später sollten solche Parser die Grundlage textbasierter Abenteuerspiele am Computer (zum Beispiel in „Mindwheel“) werden. Die Geschichte geht folgendermaßen weiter: Lustigerweise, listigerweise wählt Weizenbaum das Gebiet der Psychologie als Themenbereich, in dem die Ein- und Ausgabesätze zusammenzupassen scheinen, und es entsteht im menschlichen Gegenüber des Computers der Eindruck, ein sinnvolles Gespräch zu führen. Eine pure Illusion, denkt Weizenbaum, stellt aber fest, dass seine Sekretärin mit dem Programm spielt und Texte über intime Details in die Maschine tippt. Das war, so will es die von Weizenbaum in die Welt gesetzte Geschichte, die Initialzündung für sein großes Buch Die Macht der Computer und die Ohnmacht der Vernunft, das 1976 erschien, zehn Jahre nachdem er das Psychoprogramm geschrieben hatte (Vorüberlegungen zu diesem Buch veröffentlichte er 1972 in der ZEIT). Doch der Computerpsychiater war nicht mehr aus der Welt zu schaffen. Als „ELIZA“ ging das Programm in immer neuen Versionen um die Welt, und die computerisierte Psychotherapie ist heute sogar Gegenstand der Forschung. Das Buch. Viele hat Weizenbaum verfasst, doch das genannte Buch begründet seinen Ruhm. Es untersucht das Programmieren als Konstruieren gesellschaftlicher Wirklichkeiten. Aber eben nicht im Sinne allgemein gehaltener Kulturkritik, die mangels Detailkenntnis nicht recht zubeißen kann, sondern sehr konkret, auch auf dem Terrain der zum damaligen Zeitpunkt fortgeschrittensten Informatik. Weizenbaum kennt sie ja alle, die Künder der Künstlichen Intelligenz, er arbeitet im Massachusetts Institute of Technology Tür an Tür mit ihnen. Und während andere KI-Kritiker wie John Searle eher polemisieren anstatt sich auf die Informatik selbst einzulassen, geht Weizenbaum ins Detail, um den Teufel zu entdecken: die Vorstellung, Bedeutung könne von einer Maschine erzeugt werden. Ein folgenschwerer Irrtum, wie Weizenbaum meint, und er entfaltet auf dieser Grundlage eine Kritik der technokratischen Verschleierung von Machtverhältnissen, die zuvor von Denkern wie Herbert Marcuse oder Jürgen Habermas angedeutet, aber nicht ausgemalt werden konnte. Es folgte Weizenbaums Zeit als öffentlicher Intellektueller. Seine weiteren Bücher und Aufsätze und Reden sind, wenn man so will, Fortsetzungen dieses einen Buches. Weizenbaum intervenierte, stellte in öffentlichen Auftritten die KI-Intelligentsia bloß, ebenso die Computerpriester. Genüsslich zersägte er ihre Agitation, Propaganda und Reklame und half, in Amerika und in Deutschland Vereinigungen kritischer Informatiker zu gründen. Die Wirkungen des Macht-Ohnmacht-Buches waren weitreichend. Es trug dazu bei, dass einer der wichtigsten KI-Forscher, Terry Winograd, die Beschränktheit seiner Bemühungen erkannte und sich der Frage zuwandte, wie sich die Arbeits- und Sozialbeziehungen zwischen den Menschen mit der Informatisierung wandeln. Den Computer als Sozialmedium zu begreifen, das war der Weg, den Weizenbaum eröffnet hatte; dass später, mit der Robotik, eine andere Art Künstlicher Intelligenz auftreten sollte, die ihre eigene Körperlichkeit und sogar soziale Realität mit sich tragen würde, konnte Weizenbaum nicht voraussehen und nahm er später auch nicht ernst. Sehr erregen konnte er sich im Gespräch, gelegentlich auch mal pauschal werden, aber immer blieb er der anregende, grundfreundliche, sympathische Debattierer. Dazu ein persönliches Wort, denn „Joe“ Weizenbaum war früher oft in meinem Elternhaus zu Gast. Er hörte sich stets geduldig an, was ein computerbegeisterter Mathematiker wie mein Vater oder ein heranwachsender KI- und Vernetzungsfan wie ich so zu sagen hatten, um danach liebenswürdig - aber auf unnachahmlich ironische Art - zu antworten. Ich kenne eigentlich nur einen anderen Menschen, der es darin mit Weizenbaum aufnehmen konnte, nämlich ausgerechnet Marvin Minsky, seinen Büronachbarn, Künder der Künstlichen Intelligenz und immer der Gegenpol zu Joe. Sie konnten kultiviert miteinander sprechen. Eben das, das Gespräch als Gesellschaftsform, schien für Joseph Weizenbaum bedroht zu sein, und zwar durch explosionsartige Quatschvermehrung im Internet. Am vergangenen Mittwochabend ist Joe Weizenbaum einem Schlaganfall erlegen. Zum ThemaDIE ZEIT 11/2008:
»Heben und Rollen« DIE ZEIT 03/1972:
Albtraum Computer WHAT IS ARTIFICIAL INTELLIGENCE?John McCarthy Computer Science Department JanFebMarAprMayJun JulAugSepOctNovDec , :< 10 0 Stanford University
Revised November 24, 2004:
Abstract:
This article for the layman answers basic questions about artificial
intelligence. The opinions expressed here are not all
consensus opinion among researchers in AI.
John McCarthy 2004-11-24
AI and Computing Quotes [U]
[A] [B] [C] [D] [E] [F] [G] [H] [I] [J] [K] [L] [M] [N] [O] [P] [Q] [R] [S] [T] [U] [V] [W] [X] [Y] [Z]
History of AI
List of important publications in computer science - Wikipedia, the free encyclopedia
en.wikipedia.org/wiki/List_of_important_publicatio... Artificial intelligence[edit] Computing machinery and intelligence
Description: This paper discusses whether machine can think and suggested the Turing test as a method for checking it. In a sense, this was the beginning of artificial intelligence Importance: Topic creator, Breakthrough, Influence [edit] A Proposal for the Dartmouth Summer Research Project on Artificial IntelligenceDescription: This summer research proposal marks the areas of research in artificial intelligence since then. It was a very long summer. Importance: Influence [edit] Fuzzy sets
Description: The seminal paper published in 1965 provides details on the mathematics of fuzzy set theory. Importance: Topic creator, Influence [edit] Probabilistic Reasoning in Intelligent Systems: Networks of Plausible Inference
Description: This book introduced Bayesian methods to AI. Importance: Topic creator, Influence [edit] Artificial Intelligence: A Modern Approach
Description: The standard textbook in Artificial Intelligence. The book web site lists over 1000 colleges and universities in 93 countries using it. Importance: Introduction, Influence [edit] The Brain Makers: Genius, Ego & Greed In The Quest For Machines That ThinkDescription: The definitive book on the business of creating artificial intellligence. Importance: [edit] Machine learning[edit] Language identification in the limit
Description: This paper created Algorithmic learning theory. Importance: Topic creator, Breakthrough, Influence [edit] On the uniform convergence of relative frequencies of events to their probabilities
Description: Computational learning theory, VC theory, statistical uniform convergence and the VC dimension. Importance: Breakthrough, Influence [edit] A theory of the learnable
Description: The Probably approximately correct learning (PAC learning) framework. Importance: Topic creator, Breakthrough, Influence [edit] Learning representations by back-propagating errors
Description: Development of Backpropagation algorithm for artificial neural networks. Note that the algorithm was first described by Paul Werbos in 1974. Importance: Influence [edit] Induction of Decision Trees
Description: Decision Trees are a common learning algorithm and a decision representation tool. Development of decision trees was done by many researchers in many areas, even before this paper. Though this paper is one of the most influential in the field. Importance: Influence [edit] Learning Quickly When Irrelevant Attributes Abound: A New Linear-threshold Algorithm
Description: One of the papers that started the field of on-line learning. In this learning setting, a learner receives a sequence of examples, making predictions after each one, and receiving feedback after each prediction. Research in this area is remarkable because (1) the algorithms and proofs tend to be very simple and beautiful, and (2) the model makes no statistical assumptions about the data. In other words, the data need not be random (as in nearly all other learning models), but can be chosen arbitrarily by "nature" or even an adversary. Specifically, this paper introduced the Winnow (algorithm) algorithm. Importance: Influence [edit] Learning to predict by the method of temporal differences
Description: The temporal differences method for reinforcement learning. Importance: Influence [edit] Learnability and the Vapnik-Chervonenkis dimension
Description: The complete characterization of PAC learnability using the VC dimension. Importance: Breakthrough, Influence [edit] Cryptographic limitations on learning boolean formulae and finite automata
Description: Proving negative results for PAC learning. Importance: Influence [edit] The strength of weak learnability
Description: Proving that weak and strong learnability are equivalent in the noise free PAC framework. The proof was done by introducing the boosting method. [edit] Learning in the presence of malicious errors
Description: Proving possibility and impossibility result in the malicious errors framework. Importance: Breakthrough, Influence [edit] A training algorithm for optimum margin classifiers
Description: This paper presented support vector machines, a practical and popular machine learning algorithm. Support vector machines utilize the kernel trick, a generally used method. Importance: Breakthrough, Influence [edit] Knowledge-based analysis of microarray gene expression data by using support vector machines
Description: The first application of supervised learning to gene expression data, in particular Support Vector Machines. The method is now standard, and the paper one of the most cited in the area. Importance: Breakthrough, Influence
Intelligent Agents
Slipstream - A Software Secretary That Takes Charge - NYTimes.com
www.nytimes.com/2008/12/14/business/14stream.html?... December 14, 2008
Slipstream
A Software Secretary That Takes ChargeBy JOHN MARKOFF
SHOULDN’T your computer know a reasonable amount about your likes and dislikes? Wouldn’t it be great if it could anticipate your needs and take action without you pressing a key? Booking travel and restaurant reservations, rearranging meeting schedules or even taking a first cut at reading e-mail are among the mundane tasks that have remained beyond the reach of our PCs for decades. But now a new generation of Internet technologies, coupled with the investment of more than a third of a billion dollars, may be making meaningful progress. The concept of a software personal assistant has long captured the imagination of a generation of science fiction writers and computer scientists. Oliver G. Selfridge, the artificial-intelligence pioneer who died this month, is credited with coining the term “intelligent agent,” as well as the idea of a computer software “demon” — a simple software program that could monitor its environment and make appropriate responses when changes occur. With the arrival of personal computing in the 1980s, the idea took the form of highly choreographed “vision” statements from many Silicon Valley companies. The most memorable was the Knowledge Navigator video, by John Sculley, then chief executive of Apple, in which an interactive assistant on a video display, clad in a bow tie, does research for a college professor and nags him to return his mother’s phone call. But efforts to build useful computerized assistants have consistently ended in failure, including some of the Valley’s largest “craters” — ambitious undertakings ending as spectacular flameouts. The failures include General Magic, originally backed by Mr. Sculley, E-speak by Hewlett-Packard and Hailstorm by Microsoft. A Pentagon research project and two Silicon Valley start-up companies are about to try again. SRI International, a research group in Menlo Park, Calif., is approaching the end of a multiyear project called CALO, which stands for cognitive assistant that learns and organizes. CALO is financed by the Defense Advanced Research Projects Agency of the Pentagon and is one of the largest artificial-intelligence projects ever. Some public demonstrations have been given, but CALO is being developed largely out of the public eye because it is intended for the military. There is already one quiet commercial spin-off from the project. Siri Inc., based in San Jose, Calif., plans to introduce a personal assistance service in the first half of 2009. Still in “stealth” mode, with a small private test version of its service, Siri has raised $8.5 million from two venture capital firms. “We’re exploring concepts developed by the CALO project and applying them to the consumer,” said Adam Cheyer, Siri’s founder and vice president of engineering. He said that he expects the idea of personal assistance to gain momentum next year and that he thinks Siri will be joined by other competitors. A different tack has been taken by the entrepreneur Patrick W. Grady. He has put together a technology team at Rearden Commerce that has already begun to reach a business audience with an “intelligent” personal assistant oriented toward travel and entertainment. It will be available early next year for nonbusiness customers as well. Rearden is one of Silicon Valley’s most significantly financed but least known start-ups. Founded in 1999 before the dot-com crash, Rearden announced in April that it had recently raised an additional $100 million, for a total of $200 million in almost a decade. American Express and JPMorgan Chase each own 10 percent stakes. Why might Mr. Grady and Mr. Cheyer succeed while everyone who came before lies face down with arrows in their backs? Timing, for one thing. The promise of the Web 2.0 era of the Internet has been the interconnection of Web services. Mr. Grady says he has a far easier task today because the heavy lifting has been done by others. “This is the connective tissue that sits on top of the Web and brings you more than the sum of the parts,” he said. “I set out to deliver on the longstanding ‘holy grail of user-centric computing,’ a ‘personal Internet assistant.’” He promises to bring together all of the discrete online services needed for business travel that are now separate — for starters, travel, airport parking, car services, dining reservations, entertainment tickets, package delivery and video conferences. Imagine you are on a business trip and your computer discovers that your flight will be late. It automatically reschedules your dinner in New York, informs your three guests of the change and tells you they’ve been notified. REARDEN’S service is used by 2,500 American Express business customers, making it available to 1.6 million employees. Chase plans to start the service for its card members early next year. One of Rearden’s first customers was GlaxoSmithKline, the pharmaceutical maker. The company has made the service available to more than 50,000 users in the United States and Britain; the service helps to plan 3,000 to 4,000 trips a week. “We think of it as a personal executive assistant,” said Gregg Brandyberry, a vice p |
Article Information
Since our society places so much importance on the concept of being known, why not find out which celebrity you most resemble? Lord knows we wasted a few minutes going through our various photos, only to be told that we look like Alban Berg (who?). Using super secret face recognition algorithms, MyHeritage analyzes your facial features and compares them to those found in its large database of celebrity faces. Simply upload your mugshot to their servers and voilà, instant gratification. This is a total productivity killer, as we imagine offices everywhere scanning photos and giggling like children at the results. So if you've got a minute, why not take a look, just don't come crying to us when MyHeritage tells you that you look like Golem.
Via an 
"In 'information extraction,' computers scan text for words and phrases that identify subjects, objects and specific types of information in order to understand the highly variable ways in which human beings express themselves." (Credits: Claire Cardie (diagram) and Bill Steele (caption) for the 
Naturally Speaking 9
, you can try different tricks, and not all will work.
is a blank tile you can slide around, then you can use search to figure it out. Each state will have at most four resulting states, because you can slide 
