The Singularity Is Near_ When Humans Transcend Biology Part 13

By the late 1980s expert systems were incorporating the idea of uncertainty and could combine many sources of probabilistic evidence to make a decision. The MYCIN system pioneered this approach. A typical MYCIN "rule" reads:

If the infection which requires therapy is meningitis, and the type of the infection is fungal, and organisms were not seen on the stain of the culture, and the patient is not a compromised host, and the patient has been to an area that is endemic for coccidiomycoses, and the race of the patient is Black, Asian, or Indian, and the cryptococcal antigen in the csf test was not positive, THEN there is a 50 percent chance that cryptococcus is not one of the organisms which is causing the infection.

Although a single probabilistic rule such as this would not be sufficient by itself to make a useful statement, by combining thousands of such rules the evidence can be marshaled and combined to make reliable decisions.

Probably the longest-running expert system project is CYC (for enCYClopedic), created by Doug Lenat and his colleagues at Cycorp. Initiated in 1984, CYC has been coding commonsense knowledge to provide machines with an ability to understand the unspoken a.s.sumptions underlying human ideas and reasoning. The project has evolved from hard-coded logical rules to probabilistic ones and now includes means of extracting knowledge from written sources (with human supervision). The original goal was to generate one million rules, which reflects only a small portion of what the average human knows about the world. Lenat's latest goal is for CYC to master "100 million things, about the number a typical person knows about the world, by 2007."166 Another ambitious expert system is being pursued by Darryl Macer, a.s.sociate professor of biological sciences at the University of Tsukuba in j.a.pan. He plans to develop a system incorporating all human ideas.167 One application would be to inform policy makers of which ideas are held by which community. One application would be to inform policy makers of which ideas are held by which community.

Bayesian Nets. Over the last decade a technique called Bayesian logic has created a robust mathematical foundation for combining thousands or even millions of such probabilistic rules in what are called "belief networks" or Bayesian nets. Originally devised by English mathematician Thomas Bayes and published posthumously in 1763, the approach is intended to determine the likelihood of future events based on similar occurrences in the past. Over the last decade a technique called Bayesian logic has created a robust mathematical foundation for combining thousands or even millions of such probabilistic rules in what are called "belief networks" or Bayesian nets. Originally devised by English mathematician Thomas Bayes and published posthumously in 1763, the approach is intended to determine the likelihood of future events based on similar occurrences in the past.168 Many expert systems based on Bayesian techniques gather data from experience in an ongoing fas.h.i.+on, thereby continually learning and improving their decision making. Many expert systems based on Bayesian techniques gather data from experience in an ongoing fas.h.i.+on, thereby continually learning and improving their decision making.

The most promising type of spam filters are based on this method. I personally use a spam filter called SpamBayes, which trains itself on e-mail that you have identified as either "spam" or "okay."169 You start out by presenting a folder of each to the filter. It trains its Bayesian belief network on these two files and a.n.a.lyzes the patterns of each, thus enabling it to automatically move subsequent e-mail into the proper category. It continues to train itself on every subsequent e-mail, especially when it's corrected by the user. This filter has made the spam situation manageable for me, which is saying a lot, as it weeds out two hundred to three hundred spam messages each day, letting more than one hundred "good" messages through. Only about 1 percent of the messages it identifies as "okay" are actually spam; it almost never marks a good message as spam. The system is almost as accurate as I would be and much faster. You start out by presenting a folder of each to the filter. It trains its Bayesian belief network on these two files and a.n.a.lyzes the patterns of each, thus enabling it to automatically move subsequent e-mail into the proper category. It continues to train itself on every subsequent e-mail, especially when it's corrected by the user. This filter has made the spam situation manageable for me, which is saying a lot, as it weeds out two hundred to three hundred spam messages each day, letting more than one hundred "good" messages through. Only about 1 percent of the messages it identifies as "okay" are actually spam; it almost never marks a good message as spam. The system is almost as accurate as I would be and much faster.

Markov Models. Another method that is good at applying probabilistic networks to complex sequences of information involves Markov models. Another method that is good at applying probabilistic networks to complex sequences of information involves Markov models.170 Andrei Andreyevich Markov (18561922), a renowned mathematician, established a theory of "Markov chains," which was refined by Norbert Wiener (18941964) in 1923. The theory provided a method to evaluate the likelihood that a certain sequence of events would occur. It has been popular, for example, in speech recognition, in which the sequential events are phonemes (parts of speech). The Markov models used in speech recognition code the likelihood that specific patterns of sound are found in each phoneme, how the phonemes influence each other, and likely orders of phonemes. The system can also include probability networks on higher levels of language, such as the order of words. The actual probabilities in the models are trained on actual speech and language data, so the method is self-organizing. Andrei Andreyevich Markov (18561922), a renowned mathematician, established a theory of "Markov chains," which was refined by Norbert Wiener (18941964) in 1923. The theory provided a method to evaluate the likelihood that a certain sequence of events would occur. It has been popular, for example, in speech recognition, in which the sequential events are phonemes (parts of speech). The Markov models used in speech recognition code the likelihood that specific patterns of sound are found in each phoneme, how the phonemes influence each other, and likely orders of phonemes. The system can also include probability networks on higher levels of language, such as the order of words. The actual probabilities in the models are trained on actual speech and language data, so the method is self-organizing.

Markov modeling was one of the methods my colleagues and I used in our own speech-recognition development.171 Unlike phonetic approaches, in which specific rules about phoneme sequences are explicitly coded by human linguists, we did not tell the system that there are approximately forty-four phonemes in English, nor did we tell it what sequences of phonemes were more likely than others. We let the system discover these "rules" for itself from thousands of hours of transcribed human speech data. The advantage of this approach over hand-coded rules is that the models develop subtle probabilistic rules of which human experts are not necessarily aware. Unlike phonetic approaches, in which specific rules about phoneme sequences are explicitly coded by human linguists, we did not tell the system that there are approximately forty-four phonemes in English, nor did we tell it what sequences of phonemes were more likely than others. We let the system discover these "rules" for itself from thousands of hours of transcribed human speech data. The advantage of this approach over hand-coded rules is that the models develop subtle probabilistic rules of which human experts are not necessarily aware.

Neural Nets. Another popular self-organizing method that has also been used in speech recognition and a wide variety of other pattern-recognition tasks is neural nets. This technique involves simulating a simplified model of neurons and interneuronal connections. One basic approach to neural nets can be described as follows. Each point of a given input (for speech, each point represents two dimensions, one being frequency and the other time; for images, each point would be a pixel in a two-dimensional image) is randomly connected to the inputs of the first layer of simulated neurons. Every connection has an a.s.sociated synaptic strength, which represents its importance and which is set at a random value. Each neuron adds up the signals coming into it. If the combined signal exceeds a particular threshold, the neuron fires and sends a signal to its output connection; if the combined input signal does not exceed the threshold, the neuron does not fire, and its output is zero. The output of each neuron is randomly connected to the inputs of the neurons in the next layer. There are multiple layers (generally three or more), and the layers may be organized in a variety of configurations. For example, one layer may feed back to an earlier layer. At the top layer, the output of one or more neurons, also randomly selected, provides the answer. (For an algorithmic description of neural nets, see this note: Another popular self-organizing method that has also been used in speech recognition and a wide variety of other pattern-recognition tasks is neural nets. This technique involves simulating a simplified model of neurons and interneuronal connections. One basic approach to neural nets can be described as follows. Each point of a given input (for speech, each point represents two dimensions, one being frequency and the other time; for images, each point would be a pixel in a two-dimensional image) is randomly connected to the inputs of the first layer of simulated neurons. Every connection has an a.s.sociated synaptic strength, which represents its importance and which is set at a random value. Each neuron adds up the signals coming into it. If the combined signal exceeds a particular threshold, the neuron fires and sends a signal to its output connection; if the combined input signal does not exceed the threshold, the neuron does not fire, and its output is zero. The output of each neuron is randomly connected to the inputs of the neurons in the next layer. There are multiple layers (generally three or more), and the layers may be organized in a variety of configurations. For example, one layer may feed back to an earlier layer. At the top layer, the output of one or more neurons, also randomly selected, provides the answer. (For an algorithmic description of neural nets, see this note:172) Since the neural-net wiring and synaptic weights are initially set randomly, the answers of an untrained neural net will be random. The key to a neural net, therefore, is that it must learn its subject matter. Like the mammalian brains on which it's loosely modeled, a neural net starts out ignorant. The neural net's teacher-which may be a human, a computer program, or perhaps another, more mature neural net that has already learned its lessons-rewards the student neural net when it generates the right output and punishes it when it does not. This feedback is in turn used by the student neural net to adjust the strengths of each interneuronal connection. Connections that were consistent with the right answer are made stronger. Those that advocated a wrong answer are weakened. Over time, the neural net organizes itself to provide the right answers without coaching. Experiments have shown that neural nets can learn their subject matter even with unreliable teachers. If the teacher is correct only 60 percent of the time, the student neural net will still learn its lessons.

A powerful, well-taught neural net can emulate a wide range of human pattern-recognition faculties. Systems using multilayer neural nets have shown impressive results in a wide variety of pattern-recognition tasks, including recognizing handwriting, human faces, fraud in commercial transactions such as credit-card charges, and many others. In my own experience in using neural nets in such contexts, the most challenging engineering task is not coding the nets but in providing automated lessons for them to learn their subject matter.

The current trend in neural nets is to take advantage of more realistic and more complex models of how actual biological neural nets work, now that we are developing detailed models of neural functioning from brain reverse NEAR engineering.173 Since we do have several decades of experience in using self-organizing paradigms, new insights from brain studies can quickly be adapted to neural-net experiments. Since we do have several decades of experience in using self-organizing paradigms, new insights from brain studies can quickly be adapted to neural-net experiments.

Neural nets are also naturally amenable to parallel processing, since that is how the brain works. The human brain does not have a central processor that simulates each neuron. Rather, we can consider each neuron and each interneuronal connection to be an individual slow processor. Extensive work is under way to develop specialized chips that implement neural-net architectures in parallel to provide substantially greater throughput.174

Genetic Algorithms (GAs). Another self-organizing paradigm inspired by nature is genetic, or evolutionary, algorithms, which emulate evolution, including s.e.xual reproduction and mutations. Here is a simplified description of how they work. First, determine a way to code possible solutions to a given problem. If the problem is optimizing the design parameters for a jet engine, define a list of the parameters (with a specific number of bits a.s.signed to each parameter). This list is regarded as the genetic code in the genetic algorithm. Then randomly generate thousands or more genetic codes. Each such genetic code (which represents one set of design parameters) is considered a simulated "solution" organism. Another self-organizing paradigm inspired by nature is genetic, or evolutionary, algorithms, which emulate evolution, including s.e.xual reproduction and mutations. Here is a simplified description of how they work. First, determine a way to code possible solutions to a given problem. If the problem is optimizing the design parameters for a jet engine, define a list of the parameters (with a specific number of bits a.s.signed to each parameter). This list is regarded as the genetic code in the genetic algorithm. Then randomly generate thousands or more genetic codes. Each such genetic code (which represents one set of design parameters) is considered a simulated "solution" organism.

Now evaluate each simulated organism in a simulated environment by using a defined method to evaluate each set of parameters. This evaluation is a key to the success of a genetic algorithm. In our example, we would apply each solution organism to a jet-engine simulation and determine how successful that set of parameters is, according to whatever criteria we are interested in (fuel consumption, speed, and so on). The best solution organisms (the best designs) are allowed to survive, and the rest are eliminated.

Now have each of the survivors multiply themselves until they reach the same number of solution creatures. This is done by simulating s.e.xual reproduction. In other words, each new offspring solution draws part of its genetic code from one parent and another part from a second parent. Usually no distinction is made between male or female organisms; it's sufficient to generate an offspring from two arbitrary parents. As they multiply, allow some mutation (random change) in the chromosomes to occur.

We've now defined one generation of simulated evolution; now repeat these steps for each subsequent generation. At the end of each generation determine how much the designs have improved. When the improvement in the evaluation of the design creatures from one generation to the next becomes very small, we stop this iterative cycle of improvement and use the best design(s) in the last generation. (For an algorithmic description of genetic algorithms, see this note.175) The key to a GA is that the human designers don't directly program a solution; rather, they let one emerge through an iterative process of simulated compet.i.tion and improvement. As we discussed, biological evolution is smart but slow, so to enhance its intelligence we retain its discernment while greatly speeding up its ponderous pace. The computer is fast enough to simulate many generations in a matter of hours or days or weeks. But we have to go through this iterative process only once; once we have let this simulated evolution run its course, we can apply the evolved and highly refined rules to real problems in a rapid fas.h.i.+on.

Like neural nets GAs are a way to harness the subtle but profound patterns that exist in chaotic data. A key requirement for their success is a valid way of evaluating each possible solution. This evaluation needs to be fast because it must take account of many thousands of possible solutions for each generation of simulated evolution.

GAs are adept at handling problems with too many variables to compute precise a.n.a.lytic solutions. The design of a jet engine, for example, involves more than one hundred variables and requires satisfying dozens of constraints. GAs used by researchers at General Electric were able to come up with engine designs that met the constraints more precisely than conventional methods.

When using GAs you must, however, be careful what you ask for. University of Suss.e.x researcher Jon Bird used a GA to optimally design an oscillator circuit. Several attempts generated conventional designs using a small number of transistors, but the winning design was not an oscillator at all but a simple radio circuit. Apparently the GA discovered that the radio circuit picked up an oscillating hum from a nearby computer.176 The GA's solution worked only in the exact location on the table where it was asked to solve the problem. The GA's solution worked only in the exact location on the table where it was asked to solve the problem.

Genetic algorithms, part of the field of chaos or complexity theory, are increasingly being used to solve otherwise intractable business problems, such as optimizing complex supply chains. This approach is beginning to supplant more a.n.a.lytic methods throughout industry. (See examples below.) The paradigm is also adept at recognizing patterns, and is often combined with neural nets and other self-organizing methods. It's also a reasonable way to write computer software, particularly software that needs to find delicate balances for competing resources.

In the novel usr/bin/G.o.d usr/bin/G.o.d, Cory Doctorow, a leading science-fiction writer, uses an intriguing variation of a GA to evolve an AI. The GA generates a large number of intelligent systems based on various intricate combinations of techniques, with each combination characterized by its genetic code. These systems then evolve using a GA.

. The evaluation function works as follows: each system logs on to various human chat rooms and tries to pa.s.s for a human, basically a covert Turing test. If one of the humans in a chat room says something like "What are you, a chatterbot?" (chatterbot meaning an automatic program, which at today's level of development is expected to not understand language at a human level), the evaluation is over, that system ends its interactions, and reports its score to the GA. The score is determined by how long it was able to pa.s.s for human without being challenged in this way. The GA evolves more and more intricate combinations of techniques that are increasingly capable of pa.s.sing for human.

The main difficulty with this idea is that the evaluation function is fairly slow, although it will take an appreciable amount of time only after the systems are reasonably intelligent. Also, the evaluations can take place largely in parallel. It's an interesting idea and may actually be a useful method to finish the job of pa.s.sing the Turing test, once we get to the point where we have sufficiently sophisticated algorithms to feed into such a GA, so that evolving a Turing-capable AI is feasible.

Recursive Search. Often we need to search through a vast number of combinations of possible solutions to solve a given problem. A cla.s.sic example is in playing games such as chess. As a player considers her next move, she can list all of her possible moves, and then, for each such move, all possible countermoves by the opponent, and so on. It is difficult, however, for human players to keep a huge "tree" of move-countermove sequences in their heads, and so they rely on pattern recognition-recognizing situations based on prior experience-whereas machines use logical a.n.a.lysis of millions of moves and countermoves. Often we need to search through a vast number of combinations of possible solutions to solve a given problem. A cla.s.sic example is in playing games such as chess. As a player considers her next move, she can list all of her possible moves, and then, for each such move, all possible countermoves by the opponent, and so on. It is difficult, however, for human players to keep a huge "tree" of move-countermove sequences in their heads, and so they rely on pattern recognition-recognizing situations based on prior experience-whereas machines use logical a.n.a.lysis of millions of moves and countermoves.

Such a logical tree is at the heart of most game-playing programs. Consider how this is done. We construct a program called Pick Best Next Step to select each move. Pick Best Next Step starts by listing all of the possible moves from the current state of the board. (If the problem was solving a mathematical theorem, rather than game moves, the program would list all of the possible next steps in a proof.) For each move the program constructs a hypothetical board that reflects what would happen if we made this move. For each such hypothetical board, we now need to consider what our opponent would do if we made this move. Now recursion comes in, because Pick Best Next Step simply calls Pick Best Next Step (in other words, itself) to pick the best move for our opponent. In calling itself, Pick Best Next Step then lists all of the legal moves for our opponent.

The program keeps calling itself, looking ahead as many moves as we have time to consider, which results in the generation of a huge move-countermove tree. This is another example of exponential growth, because to look ahead an additional move (or countermove) requires multiplying the amount of available computation by about five. Key to the success of the recursive formula is pruning this huge tree of possibilities and ultimately stopping its growth. In the game context, if a board looks hopeless for either side, the program can stop the expansion of the move-countermove tree from that point (called a "terminal leaf" of the tree) and consider the most recently considered move to be a likely win or loss. When all of these nested program calls are completed, the program will have determined the best possible move for the current actual board within the limits of the depth of recursive expansion that it had time to pursue and the quality of its pruning algorithm. (For an algorithmic description of recursive search, see this note:177) The recursive formula is often effective at mathematics. Rather than game moves, the "moves" are the axioms of the field of math being addressed, as well as previously proved theorems. The expansion at each point is the possible axioms (or previously proved theorems) that can be applied to a proof at each step. (This was the approach used by Newell, Shaw, and Simons's General Problem Solver.) From these examples it may appear that recursion is well suited only for problems in which we have crisply defined rules and objectives. But it has also shown promise in computer generation of artistic creations. For example, a program I designed called Ray Kurzweil's Cybernetic Poet uses a recursive approach.178 The program establishes a set of goals for each word-achieving a certain rhythmic pattern, poem structure, and word choice that is desirable at that point in the poem. If the program is unable to find a word that meets these criteria, it backs up and erases the previous word it has written, reestablishes the criteria it had originally set for the word just erased, and goes from there. If that also leads to a dead end, it backs up again, thus moving backward and forward. Eventually, it forces itself to make up its mind by relaxing some of the constraints if all paths lead to dead ends. The program establishes a set of goals for each word-achieving a certain rhythmic pattern, poem structure, and word choice that is desirable at that point in the poem. If the program is unable to find a word that meets these criteria, it backs up and erases the previous word it has written, reestablishes the criteria it had originally set for the word just erased, and goes from there. If that also leads to a dead end, it backs up again, thus moving backward and forward. Eventually, it forces itself to make up its mind by relaxing some of the constraints if all paths lead to dead ends.

Combining Methods. The most powerful approach to building robust AI systems is to combine approaches, which is how the human brain works. As we discussed, the brain is not one big neural net but instead consists of hundreds of regions, each of which is optimized for processing information in a different way. None of these regions by itself operates at what we would consider human levels of performance, but dearly by definition the overall system does exactly that. The most powerful approach to building robust AI systems is to combine approaches, which is how the human brain works. As we discussed, the brain is not one big neural net but instead consists of hundreds of regions, each of which is optimized for processing information in a different way. None of these regions by itself operates at what we would consider human levels of performance, but dearly by definition the overall system does exactly that.

I've used this approach in my own AI work, especially in pattern recognition. In speech recognition, for example, we implemented a number of different pattern-recognition systems based on different paradigms. Some were specifically programmed with knowledge of phonetic and linguistic constraints from experts. Some were based on rules to pa.r.s.e sentences (which involves creating sentence diagrams showing word usage, similar to the diagrams taught in grade school). Some were based on self-organizing techniques, such as Markov models, trained on extensive libraries of recorded and annotated human speech. We then programmed a software "expert manager" to learn the strengths and weaknesses of the different "experts" (recognizers) and combine their results in optimal ways. In this fas.h.i.+on, a particular technique that by itself might produce unreliable results can nonetheless contribute to increasing the overall accuracy of the system.

There are many intricate ways to combine the varied methods in AI's toolbox. For example, one can use a genetic algorithm to evolve the optimal topology (organization of nodes and connections) for a neural net or a Markov model. The final output of the GA-evolved neural net can then be used to control the parameters of a recursive search algorithm. We can add in powerful signal- and image-processing techniques that have been developed for pattern-processing systems. Each specific application calls for a different architecture. Computer science professor and AI entrepreneur Ben Goertzel has written a series of books and articles that describe strategies and architectures for combining the diverse methods underlying intelligence. His Novamente architecture is intended to provide a framework for general-purpose AI.179 The above basic descriptions provide only a glimpse into how increasingly sophisticated current AI systems are designed. It's beyond the scope of this book to provide a comprehensive description of the techniques of AI, and even a doctoral program in computer science is unable to cover all of the varied approaches in use today.

Many of the examples of real-world narrow AI systems described in the next section use a variety of methods integrated and optimized for each particular task. Narrow AI is strengthening as a result of several concurrent trends: continued exponential gains in computational resources, extensive real-world experience with thousands of applications, and fresh insights into how the human brain makes intelligent decisions.

A Narrow AI Sampler

When I wrote my first AI book, The Age of Intelligent Machines, in the late 1980s, I had to conduct extensive investigations to find a few successful examples of AI in practice. The Internet was not yet prevalent, so I had to go to real libraries and visit the AI research centers in the United States, Europe, and Asia. I included in my book pretty much all of the reasonable examples I could identify. In my research for this book my experience has been altogether different. I have been inundated with thousands of compelling examples. In our reporting on the KurzweilAI.net Web site, we feature one or more dramatic systems almost every day.180 A 2003 study by Business Communications Company projected a $21 billion market by 2007 for AI applications, with average annual growth of 12.2 percent from 2002 to 2007.181 Leading industries for AI applications include business intelligence, customer relations, finance, defense and domestic security, and education. Here is a small sample of narrow AI in action. Leading industries for AI applications include business intelligence, customer relations, finance, defense and domestic security, and education. Here is a small sample of narrow AI in action.

Military and Intelligence. The U.S. military has been an avid user of AI systems. Pattern-recognition software systems guide autonomous weapons such as cruise missiles, which can fly thousands of miles to find a specific building or even a specific window. The U.S. military has been an avid user of AI systems. Pattern-recognition software systems guide autonomous weapons such as cruise missiles, which can fly thousands of miles to find a specific building or even a specific window.182 Although the relevant details of the terrain that the missile flies over are programmed ahead of time, variations in weather, ground cover, and other factors require a flexible level of real-time image recognition. Although the relevant details of the terrain that the missile flies over are programmed ahead of time, variations in weather, ground cover, and other factors require a flexible level of real-time image recognition.

The army has developed prototypes of self-organizing communication networks (called "mesh networks") to automatically configure many thousands of communication nodes when a platoon is dropped into a new location.183 Expert systems incorporating Bayesian networks and GAs are used to optimize complex supply chains that coordinate millions of provisions, supplies, and weapons based on rapidly changing battlefield requirements.

AI systems are routinely employed to simulate the performance of weapons, including nuclear bombs and missiles.

Advance warning of the September 11, 2001, terrorist attacks was apparently detected by the National Security Agency's AI-based Echelon system, which a.n.a.lyzes the agency's extensive monitoring of communications traffic.184 Unfortunately, Echelon's warnings were not reviewed by human agents until it was too late. Unfortunately, Echelon's warnings were not reviewed by human agents until it was too late.

The 2002 military campaign in Afghanistan saw the debut of the armed Predator, an unmanned robotic flying fighter. Although the air force's Predator had been under development for many years, arming it with army-supplied missiles was a last-minute improvisation that proved remarkably successful. In the Iraq war that began in 2003 the armed Predator (operated by the CIA) and other flying unmanned aerial vehicles (UAVs) destroyed thousands of enemy tanks and missile sites.

All of the military services are using robots. The army utilizes them to search caves (in Afghanistan) and buildings. The navy uses small robotic s.h.i.+ps to protect its aircraft carriers. As I discuss in the next chapter, moving soldiers away from battle is a rapidly growing trend.

s.p.a.ce Exploration. NASA is building self-understanding into the software controlling its unmanned s.p.a.cecraft. Because Mars is about three light-minutes from Earth, and Jupiter around forty light-minutes (depending on the exact position of the planets), communication between s.p.a.cecraft headed there and earthbound controllers is significantly delayed. For this reason it's important that the software controlling these missions have the capability of performing its own tactical decision making. To accomplish this NASA software is being designed to include a model of the software's own capabilities and those of the s.p.a.cecraft, as well as the challenges each mission is likely to encounter. Such AI-based systems are capable of reasoning through new situations rather than just following preprogrammed rules. This approach enabled the craft Deep s.p.a.ce ** One in 1999 to use its own technical knowledge to devise a series of original plans to overcome a stuck switch that threatened to destroy its mission of exploring an asteroid. NASA is building self-understanding into the software controlling its unmanned s.p.a.cecraft. Because Mars is about three light-minutes from Earth, and Jupiter around forty light-minutes (depending on the exact position of the planets), communication between s.p.a.cecraft headed there and earthbound controllers is significantly delayed. For this reason it's important that the software controlling these missions have the capability of performing its own tactical decision making. To accomplish this NASA software is being designed to include a model of the software's own capabilities and those of the s.p.a.cecraft, as well as the challenges each mission is likely to encounter. Such AI-based systems are capable of reasoning through new situations rather than just following preprogrammed rules. This approach enabled the craft Deep s.p.a.ce ** One in 1999 to use its own technical knowledge to devise a series of original plans to overcome a stuck switch that threatened to destroy its mission of exploring an asteroid.185 The AI system's first plan failed to work, but its second plan saved the mission. "These systems have a commonsense model of the physics of their internal components," explains Brian Williams, coinventor of Deep s.p.a.ce One's autonomous software and now a scientist at MIT's s.p.a.ce Systems and AI laboratories. "[The s.p.a.cecraft] can reason from that model to determine what is wrong and to know how to act." The AI system's first plan failed to work, but its second plan saved the mission. "These systems have a commonsense model of the physics of their internal components," explains Brian Williams, coinventor of Deep s.p.a.ce One's autonomous software and now a scientist at MIT's s.p.a.ce Systems and AI laboratories. "[The s.p.a.cecraft] can reason from that model to determine what is wrong and to know how to act."

Using a network of computers NASA used GAs to evolve an antenna design for three s.p.a.ce Technology 5 satellites that will study the Earth's magnetic field. Millions of possible designs competed in the simulated evolution. According to NASA scientist and project leader Jason Lohn, "We are now using the [GA] software to design tiny microscopic machines, including gyroscopes, for s.p.a.ceflight navigation. The software also may invent designs that no human designer would ever think of."186 Another NASA AI system learned on its own to distinguish stars from galaxies in very faint images with an accuracy surpa.s.sing that of human astronomers.

New land-based robotic telescopes are able to make their own decisions on where to look and how to optimize the likelihood of finding desired phenomena. Called "autonomous, semi-intelligent observatories," the systems can adjust to the weather, notice items of interest, and decide on their own to track them. They are able to detect very subtle phenomena, such as a star blinking for a nanosecond, which may indicate a small asteroid in the outer regions of our solar system pa.s.sing in front of the light from that star.187 One such system, called Moving Object and Transient Event Search System (MOTESS), has identified on its own 180 new asteroids and several comets during its first two years of operation. "We have an intelligent observing system," explained University of Exeter astronomer Alasdair Allan. "It thinks and reacts for itself, deciding whether something it has discovered is interesting enough to need more observations. If more observations are needed, it just goes ahead and gets them." One such system, called Moving Object and Transient Event Search System (MOTESS), has identified on its own 180 new asteroids and several comets during its first two years of operation. "We have an intelligent observing system," explained University of Exeter astronomer Alasdair Allan. "It thinks and reacts for itself, deciding whether something it has discovered is interesting enough to need more observations. If more observations are needed, it just goes ahead and gets them."

Similar systems are used by the military to automatically a.n.a.lyze data from spy satellites. Current satellite technology is capable of observing ground-level features about an inch in size and is not affected by bad weather, clouds, or darkness.188 The ma.s.sive amount of data continually generated would not be manageable without automated image recognition programmed to look for relevant developments. The ma.s.sive amount of data continually generated would not be manageable without automated image recognition programmed to look for relevant developments.

Medicine. If you obtain an electrocardiogram (ECG) your doctor is likely to receive an automated diagnosis using pattern recognition applied to ECG recordings. My own company (Kurzweil Technologies) is working with United Therapeutics to develop a new generation of automated ECG a.n.a.lysis for long-term un.o.btrusive monitoring (via sensors embedded in clothing and wireless communication using a cell phone) of the early warning signs of heart disease. If you obtain an electrocardiogram (ECG) your doctor is likely to receive an automated diagnosis using pattern recognition applied to ECG recordings. My own company (Kurzweil Technologies) is working with United Therapeutics to develop a new generation of automated ECG a.n.a.lysis for long-term un.o.btrusive monitoring (via sensors embedded in clothing and wireless communication using a cell phone) of the early warning signs of heart disease.189 Other pattern-recognition systems are used to diagnose a variety of imaging data. Other pattern-recognition systems are used to diagnose a variety of imaging data.

Every major drug developer is using AI programs to do pattern recognition and intelligent data mining in the development of new drug therapies. For example SRI International is building flexible knowledge bases that encode everything we know about a dozen disease agents, including tuberculosis and H. pylori (the bacteria that cause ulcers).190 The goal is to apply intelligent datamining tools (software that can search for new relations.h.i.+ps in data) to find new ways to kill or disrupt the metabolisms of these pathogens. The goal is to apply intelligent datamining tools (software that can search for new relations.h.i.+ps in data) to find new ways to kill or disrupt the metabolisms of these pathogens.

Similar systems are being applied to performing the automatic discovery of new therapies for other diseases, as well as understanding the function of genes and their roles in disease.191 For example Abbott Laboratories claims that six human researchers in one of its new labs equipped with AI-based robotic and data-a.n.a.lysis systems are able to match the results of two hundred scientists in its older drug-development labs. For example Abbott Laboratories claims that six human researchers in one of its new labs equipped with AI-based robotic and data-a.n.a.lysis systems are able to match the results of two hundred scientists in its older drug-development labs.192 Men with elevated prostate-specific antigen (PSA) levels typically undergo surgical biopsy, but about 75 percent of these men do not have prostate cancer. A new test, based on pattern recognition of proteins in the blood, would reduce this false positive rate to about 29 percent.193 The test is based on an AI program designed by Correlogic Systems in Bethesda, Maryland, and the accuracy is expected to improve further with continued development. The test is based on an AI program designed by Correlogic Systems in Bethesda, Maryland, and the accuracy is expected to improve further with continued development.

Pattern recognition applied to protein patterns has also been used in the detection of ovarian cancer. The best contemporary test for ovarian cancer, called CA-125, employed in combination with ultrasound, misses almost all early-stage tumors. "By the time it is now diagnosed, ovarian cancer is too often deadly," says Emanuel Petricoin III, codirector of the Clinical Proteomics Program run by the FDA and the National Cancer Inst.i.tute. Petricoin is the lead developer of a new AI-based test looking for unique patterns of proteins found only in the presence of cancer. In an evaluation involving hundreds of blood samples, the test was, according to Petricoin, "an astonis.h.i.+ng 100% accurate in detecting cancer, even at the earliest stages."194 About 10 percent of all Pap-smear slides in the United States are a.n.a.lyzed by a self-learning AI program called FocalPoint, developed by TriPath Imaging. The developers started out by interviewing pathologists on the criteria they use. The AI system then continued to learn by watching expert pathologists. Only the best human diagnosticians were allowed to be observed by the program. "That's the advantage of an expert system," explains Bob Schmidt, Tri-Path's technical product manager. "It allows you to replicate your very best people."

Ohio State University Health System has developed a computerized physician order-entry (CPOE) system based on an expert system with extensive knowledge across multiple specialties.195 The system automatically checks every order for possible allergies in the patient, drug interactions, duplications, drug restrictions, dosing guidelines, and appropriateness given information about the patient from the hospital's laboratory and radiology departments. The system automatically checks every order for possible allergies in the patient, drug interactions, duplications, drug restrictions, dosing guidelines, and appropriateness given information about the patient from the hospital's laboratory and radiology departments.

Science and Math. A "robot scientist" has been developed at the University of Wales that combines an AI-based system capable of formulating original theories, a robotic system that can automatically carry out experiments, and a reasoning engine to evaluate results. The researchers provided their creation with a model of gene expression in yeast. The system "automatically originates hypotheses to explain observations, devises experiments to test these hypotheses, physically runs the experiments using a laboratory robot, interprets the results to falsify hypotheses inconsistent with the data, and then repeats the cycle." A "robot scientist" has been developed at the University of Wales that combines an AI-based system capable of formulating original theories, a robotic system that can automatically carry out experiments, and a reasoning engine to evaluate results. The researchers provided their creation with a model of gene expression in yeast. The system "automatically originates hypotheses to explain observations, devises experiments to test these hypotheses, physically runs the experiments using a laboratory robot, interprets the results to falsify hypotheses inconsistent with the data, and then repeats the cycle."196 The system is capable of improving its performance by learning from its own experience. The experiments designed by the robot scientist were three times less expensive than those designed by human scientists. A test of the machine against a group of human scientists showed that the discoveries made by the machine were comparable to those made by the humans. The system is capable of improving its performance by learning from its own experience. The experiments designed by the robot scientist were three times less expensive than those designed by human scientists. A test of the machine against a group of human scientists showed that the discoveries made by the machine were comparable to those made by the humans.

Mike Young, director of biology at the University of Wales, was one of the human scientists who lost to the machine. He explains that "the robot did beat me, but only because I hit the wrong key at one point."

A long-standing conjecture in algebra was finally proved by an AI system at Argonne National Laboratory. Human mathematicians called the proof "creative."

Business, Finance, and Manufacturing. Companies in every industry are using AI systems to control and optimize logistics, detect fraud and money laundering, and perform intelligent data mining on the horde of information they gather each day. Wal-Mart, for example, gathers vast amounts of information from its transactions with shoppers. AI-based tools using neural nets and expert systems review this data to provide market-research reports for managers. This intelligent data mining allows them to make remarkably accurate predictions of the inventory required for each product in each store for each day. Companies in every industry are using AI systems to control and optimize logistics, detect fraud and money laundering, and perform intelligent data mining on the horde of information they gather each day. Wal-Mart, for example, gathers vast amounts of information from its transactions with shoppers. AI-based tools using neural nets and expert systems review this data to provide market-research reports for managers. This intelligent data mining allows them to make remarkably accurate predictions of the inventory required for each product in each store for each day.197 AI-based programs are routinely used to detect fraud in financial transactions. Future Route, an English company, for example, offers iHex, based on AI routines developed at Oxford University, to detect fraud in credit-card transactions and loan applications.198 The system continuously generates and updates its own rules based on its experience. First Union Home Equity Bank in Charlotte, North Carolina, uses Loan Arranger, a similar AI-based system, to decide whether to approve mortgage applications. The system continuously generates and updates its own rules based on its experience. First Union Home Equity Bank in Charlotte, North Carolina, uses Loan Arranger, a similar AI-based system, to decide whether to approve mortgage applications.199 NASDAQ similarly uses a learning program called the Securities Observation, News a.n.a.lysis, and Regulation (SONAR) system to monitor all trades for fraud as well as the possibility of insider trading.200 As of the end of 2003 more than 180 incidents had been detected by SONAR and referred to the U.S. Securities and Exchange Commission and Department of Justice. These included several cases that later received significant news coverage. As of the end of 2003 more than 180 incidents had been detected by SONAR and referred to the U.S. Securities and Exchange Commission and Department of Justice. These included several cases that later received significant news coverage.

Ascent Technology, founded by Patrick Winston, who directed MIT's AI Lab from 1972 through 1997, has designed a GA-based system called Smart-Airport Operations Center (SAOe) that can optimize the complex logistics of an airport, such as balancing work a.s.signments of hundreds of employees, making gate and equipment a.s.signments, and managing a myriad of other details.201 Winston points out that "figuring out ways to optimize a complicated situation is what genetic algorithms do." SAOC has raised productivity by approximately 30 percent in the airports where it has been implemented. Winston points out that "figuring out ways to optimize a complicated situation is what genetic algorithms do." SAOC has raised productivity by approximately 30 percent in the airports where it has been implemented.

Ascent's first contract was to apply its AI techniques to managing the logistics for the 1991 Desert Storm campaign in Iraq. DARPA claimed that Al-based logistic-planning systems, including the Ascent system, resulted in more savings than the entire government research investment in AI over several decades.

A recent trend in software is for AI systems to monitor a complex software system's performance, recognize malfunctions, and determine the best way to recover automatically without necessarily informing the human user.202 The idea stems from the realization that as software systems become more complex, like humans, they will never be perfect, and that eliminating all bugs is impossible. As humans, we use the same strategy: we don't expect to be perfect, but we usually try to recover from inevitable mistakes. "We want to stand this notion of systems management on its head," says Armando Fox, the head of Stanford University's Software Infrastructures Group, who is working on what is now called "autonomic computing." Fox adds, "The system has to be able to set itself up, it has to optimize itself. It has to repair itself, and if something goes wrong, it has to know how to respond to external threats." IBM, Microsoft, and other software vendors are all developing systems that incorporate autonomic capabilities. The idea stems from the realization that as software systems become more complex, like humans, they will never be perfect, and that eliminating all bugs is impossible. As humans, we use the same strategy: we don't expect to be perfect, but we usually try to recover from inevitable mistakes. "We want to stand this notion of systems management on its head," says Armando Fox, the head of Stanford University's Software Infrastructures Group, who is working on what is now called "autonomic computing." Fox adds, "The system has to be able to set itself up, it has to optimize itself. It has to repair itself, and if something goes wrong, it has to know how to respond to external threats." IBM, Microsoft, and other software vendors are all developing systems that incorporate autonomic capabilities.

Manufacturing and Robotics. Computer-integrated manufacturing (CIM) increasingly employs AI techniques to optimize the use of resources, streamline logistics, and reduce inventories through just-in-time purchasing of parts and supplies. A new trend in CIM systems is to use "case-based reasoning" rather than hard-coded, rule-based expert systems. Such reasoning codes knowledge as "cases," which are examples of problems with solutions. Initial cases are usually designed by the engineers, but the key to a successful case-based reasoning system is its ability to gather new cases from actual experience. The system is then able to apply the reasoning from its stored cases to new situations. Computer-integrated manufacturing (CIM) increasingly employs AI techniques to optimize the use of resources, streamline logistics, and reduce inventories through just-in-time purchasing of parts and supplies. A new trend in CIM systems is to use "case-based reasoning" rather than hard-coded, rule-based expert systems. Such reasoning codes knowledge as "cases," which are examples of problems with solutions. Initial cases are usually designed by the engineers, but the key to a successful case-based reasoning system is its ability to gather new cases from actual experience. The system is then able to apply the reasoning from its stored cases to new situations.

Robots are extensively used in manufacturing. The latest generation of robots uses flexible Al-based machine-vision systems-from companies such as Cognex Corporation in Natick, Ma.s.sachusetts-that can respond flexibly to varying conditions. This reduces the need for precise setup for the robot to operate correctly. Brian Carlisle, CEO of Adept Technologies, a Livermore, California, factory-automation company, points out that "even if labor costs were eliminated [as a consideration], a strong case can still be made for automating with robots and other flexible automation. In addition to quality and throughput, users gain by enabling rapid product changeover and evolution that can't be matched with hard tooling."

One of AI's leading roboticists, Hans Moravec, has founded a company called Seegrid to apply his machine-vision technology to applications in manufacturing, materials handling, and military missions.203 Moravec's software enables a device (a robot or just a material-handling cart) to walk or roll through an unstructured environment and in a single pa.s.s build a reliable "voxel" (three-dimensional pixel) map of the environment. The robot can then use the map and its own reasoning ability to determine an optimal and obstacle-free path to carry out its a.s.signed mission. Moravec's software enables a device (a robot or just a material-handling cart) to walk or roll through an unstructured environment and in a single pa.s.s build a reliable "voxel" (three-dimensional pixel) map of the environment. The robot can then use the map and its own reasoning ability to determine an optimal and obstacle-free path to carry out its a.s.signed mission.

This technology enables autonomous carts to transfer materials throughout a manufacturing process without the high degree of preparation required with conventional preprogrammed robotic systems. In military situations autonomous vehicles could carry out precise missions while adjusting to rapidly changing environments and battlefield conditions.

Machine vision is also improving the ability of robots to interact with humans. Using small, inexpensive cameras, head- and eye-tracking software can sense where a human user is, allowing robots, as well as virtual personalities on a screen, to maintain eye contact, a key element for natural interactions. Head- and eye-tracking systems have been developed at Carnegie Mellon University and MIT and are offered by small companies such as Seeing Machines of Australia.

An impressive demonstration of machine vision was a vehicle that was driven by an AI system with no human intervention for almost the entire distance from Was.h.i.+ngton, D.C., to San Diego.204 Bruce Buchanan, computer-science professor at the University of Pittsburgh and president of the American a.s.sociation of Artificial Intelligence, pointed out that this feat would have been "unheard of 10 years ago." Bruce Buchanan, computer-science professor at the University of Pittsburgh and president of the American a.s.sociation of Artificial Intelligence, pointed out that this feat would have been "unheard of 10 years ago."

Palo Alto Research Center (PARC) is developing a swarm of robots that can navigate in complex environments, such as a disaster zone, and find items of interest, such as humans who may be injured. In a September 2004 demonstration at an AI conference in San Jose, they demonstrated a group of self-organizing robots on a mock but realistic disaster area.205 The robots moved over the rough terrain, communicated with one another, used pattern recognition on images, and detected body heat to locate humans. The robots moved over the rough terrain, communicated with one another, used pattern recognition on images, and detected body heat to locate humans.

Speech and Language. Dealing naturally with language is the most challenging task of all for artificial intelligence. No simple tricks, short of fully mastering the principles of human intelligence, will allow a computerized system to convincingly emulate human conversation, even if restricted to just text messages. This was Turing's enduring insight in designing his eponymous test based entirely on written language. Dealing naturally with language is the most challenging task of all for artificial intelligence. No simple tricks, short of fully mastering the principles of human intelligence, will allow a computerized system to convincingly emulate human conversation, even if restricted to just text messages. This was Turing's enduring insight in designing his eponymous test based entirely on written language.

Although not yet at human levels, natural language-processing systems are making solid progress. Search engines have become so popular that "Google" has gone from a proper noun to a common verb, and its technology has revolutionized research and access to knowledge. Google and other search engines use Al-based statistical-learning methods and logical inference to determine the ranking of links. The most obvious failing of these search engines is their inability to understand the context of words. Although an experienced user learns how to design a string of keywords to find the most relevant sites (for example, a search for "computer chip" is likely to avoid references to potato chips that a search for "chip" alone might turn up), what we would really like to be able to do is converse with our search engines in natural language. Microsoft has developed a natural-language search engine called Ask MSR (Ask Micro-Soft Research), which actually answers natural-language questions such as "When was Mickey Mantle born?"206 After the system pa.r.s.es the sentence to determine the parts of speech (subject, verb, object, adjective and adverb modifiers, and so on), a special search engine then finds matches based on the pa.r.s.ed sentence. The found doc.u.ments are searched for sentences that appear to answer the question, and the possible answers are ranked. At least 75 percent of the time, the correct answer is in the top three ranked positions, and incorrect answers are usually obvious (such as "Mickey Mantle was born in 3"). The researchers hope to include knowledge bases that will lower the rank of many of the nonsensical answers. After the system pa.r.s.es the sentence to determine the parts of speech (subject, verb, object, adjective and adverb modifiers, and so on), a special search engine then finds matches based on the pa.r.s.ed sentence. The found doc.u.ments are searched for sentences that appear to answer the question, and the possible answers are ranked. At least 75 percent of the time, the correct answer is in the top three ranked positions, and incorrect answers are usually obvious (such as "Mickey Mantle was born in 3"). The researchers hope to include knowledge bases that will lower the rank of many of the nonsensical answers.

Microsoft researcher Eric Brill, who has led research on Ask MSR, has also attempted an even more difficult task: building a system that provides answers of about fifty words to more complex questions, such as, "How are the recipients of the n.o.bel Prize selected?" One of the strategies used by this system is to find an appropriate FAQ section on the Web that answers the query.

Natural-language systems combined with large-vocabulary, speaker-independent (that is, responsive to any speaker) speech recognition over the phone are entering the marketplace to conduct routine transactions. You can talk to British Airways' virtual travel agent about anything you like as long as it has to do with booking flights on British Airways.207 You're also likely to talk to a virtual person if you call Verizon for customer service or Charles Schwab and Merrill Lynch to conduct financial transactions. These systems, while they can be annoying to some people, are reasonably adept at responding appropriately to the often ambiguous and fragmented way people speak. Microsoft and other companies are offering systems that allow a business to create virtual agents to book reservations for travel and hotels and conduct routine transactions of all kinds through two-way, reasonably natural voice dialogues. You're also likely to talk to a virtual person if you call Verizon for customer service or Charles Schwab and Merrill Lynch to conduct financial transactions. These systems, while they can be annoying to some people, are reasonably adept at responding appropriately to the often ambiguous and fragmented way people speak. Microsoft and other companies are offering systems that allow a business to create virtual agents to book reservations for travel and hotels and conduct routine transactions of all kinds through two-way, reasonably natural voice dialogues.

Not every caller is satisfied with the ability of these virtual agents to get the job done, but most systems provide a means to get a human on the line. Companies using these systems report that they reduce the need for human service agents up to 80 percent. Aside from the money saved, reducing the size of call centers has a management benefit. Call-center jobs have very high turnover rates because of low job satisfaction.

It's said that men are loath to ask others for directions, but car vendors are betting that both male and female drivers will be willing to ask their own car for help in getting to their destination. In 2005 the Acura RL and Honda Odyssey will be offering a system from IBM that allows users to converse with their cars.208 Driving directions will include street names (for example, "turn left on Main Street, then right on Second Avenue"). Users can ask such questions as "Where is the nearest Italian restaurant?" or they can enter specific locations by voice, ask for clarifications on directions, and give commands to the car itself (such as "Turn up the air conditioning"). The Acura RL will also track road conditions and highlight traffic congestion on its screen in real time. Driving directions will include street names (for example, "turn left on Main Street, then right on Second Avenue"). Users can ask such questions as "Where is the nearest Italian restaurant?" or they can enter specific locations by voice, ask for clarifications on directions, and give commands to the car itself (such as "Turn up the air conditioning"). The Acura RL will also track road conditions and highlight traffic congestion on its screen in real time.

The speech recognition is claimed to be speaker-independent and to be unaffected by engine sound, wind, and other noises. The system will reportedly recognize 1.7 million street and city names, in addition to nearly one thousand commands.

Computer language translation continues to improve gradually. Because this is a Turing-level task-that is, it requires full human-level understanding of language to perform at human levels-it will be one of the last application areas to compete with human performance. Franz Josef Och, a computer scientist at the University of Southern California, has developed a technique that can generate a new language-translation system between any pair of languages in a matter of hours or days.209 All he needs is a "Rosetta stone"-that is, text in one language and the translation of that text in the other language-although he needs millions of words of such translated text. Using a self-organizing technique, the system is able to develop its own statistical models of how text is translated from one language to the other and develops these models in both directions. All he needs is a "Rosetta stone"-that is, text in one language and the translation of that text in the other language-although he needs millions of words of such translated text. Using a self-organizing technique, the system is able to develop its own statistical models of how text is translated from one language to the other and develops these models in both directions.

This contrasts with other translation systems, in which linguists painstakingly code grammar rules with long lists of exceptions to each rule. Och's system recently received the highest score in a compet.i.tion of translation systems conducted by the U.S. Commerce Department's National Inst.i.tute of Standards and Technology.