The Emergence of Intelligent Agents:
Perhaps encouraged by the progress in solving the subproblems of AI, researchers have also started to look at the “whole agent” problem again. The work of Allen Newell, John Laird, and Paul Rosenbloom on SOAR (Newell, 1990; Laird et al., 1987) is the best-known example of a complete agent architecture. The so-called situated movement aims to understand the workings of agents embedded for intelligent agent is the Internet. AI systems have become so common in web based applications that the “-bot” suffix has entered everyday language. Moreover, AI technologies underlie many Internet tools, such as search engines, recommender systems, and web site construction systems.
Besides the first edition of this text (Russell and Norvig, 1955), other recent texts have also adopted the agent perspective (Poole et al., 1998; Nilsson, 1998). One consequence of trying to build complete agents is the realization that the previously isolated subfields of AI might need to be reorganized somewhat when their results are to be tied together. In particular, it is now widely appreciated that sensory systems (vision, sonar, speech recognition, etc.) cannot deliver perfectly reliable information about the environment. Hence, reasoning and planning systems must be able to handle uncertainly. A second major consequence of the agent perspective is that AI has been drawn into much closer contact with other fields, such as control theory and economics that also deal with agents.
Application Domains:
Industry and Agriculture:
Traditionally, robots have been fielded in areas that require difficult human labor, yet are structured enough to be amenable to robotic automation. The best example is the assembly line, where manipulators routinely perform tasks such as assembly, part placement, material handling, welding, and planning. In many of these tasks, robots have become more cost-effective than human workers.
Outdoors, many of the heavy machines that we use to harvest, mine, or excavate earth have been turned into robots. For example, a recent project at Carnegie Mellon has demonstrated that robots can strip paint off large ships about 50 times faster than people can, and with a much reduced environment impact. Prototypes of autonomous mining robots have been found to be faster and more precise than people in transporting ore in underground mines. Robots have been used to generate high-precision maps of abandoned mines and sewer systems. While many of these systems are still in their prototype stages, it is only a matter of time until robots will take over much of the semi-mechanical work that is presently performed by people.
Transportation:
Robotic transportation has many facets: from autonomous helicopters that deliver objects to locations that would be hard to access by other means, to automatic wheelchairs that transport people who are unable to control wheelchairs by themselves, to autonomous straddle carriers that outperform skilled human drivers when transporting containers from ships to trucks on loading docks. Researchers have developed car-like robotic systems that can navigate autonomous vehicles are row routinely deployed to transport goods in warehouses and between production lines.
Many of these robots require environmental modifications for their operation. The most common modifications are localization aids such as inductive loops in the floor, active beacons, barcodes tags, and GPS satellite. An open challenge in robotics is the design of robots that can use natural cues, instead of artificial devices, to navigate, particularly in environments such as the deep ocean where GPS is unavailable.
Hazardous environments:
Robots have assisted people in cleaning up nuclear waste, most notably in Chernobyl and Three Mile Island. Robots were present after the collapse of the World Trade Center, where they entered structures deemed too dangerous for human search and rescue crews.
Some countries have used robots to transport ammunition and to defuse bombs-a notoriously dangerous task. A number of research projects are presently developing prototype robots for clearing minefields, on land and at sea. Most existing robots for these tasks are teleoperated-a human operates them by remote control. Providing such robots with autonomy is an important next step.
Exploration:
Robots have gone where no-one has gone before, including the surface of Mars. Robotic arms assist astronauts in deploying and retrieving satellite and in building the International Space Station. Robots also help explore under the sea. They are routinely used to acquire maps of sunken ships. Unmanned air vehicles known as drones are used in militar6y operations. Robots are becoming very effective tools for gathering information in domains that are difficult to access for people.
Health care:
Robots are increasingly used to assist surgeons with instrument placement when operating on organs as intricate as brains, eyes, and hearts. Robots have become indispensable tools in certain types of hip replacements, thanks to their high precision. In pilot studies, robotic devices have been found to reduce the danger of lesions when performing colonoscopies. Outside the operating room, researchers have begun to develop robotic aides for elderly and handicapped people, such as intelligent robotic walkers and intelligent toys that provide reminders to take medication.
Personal services:
` Service is an up and coming application domain of robotics. Service robots assist individuals in performing daily tasks. Commercially available domestic service robots include autonomous vacuum cleaners, lawn mowers, and golf caddies. All these robots can navigate autonomously and perform their tasks without human help. Some service robots operate in public places, such as robotic information kiosks that have been deployed in shopping malls and trade fairs, or in museums as tour guides. Service tasks require human interaction, and the ability to cope robustly with unpredictable and dynamic environments.
Entertainment:
Robots have begun to conquer the entertainment and toy industry. We saw the Sony AIBO. One of the challenging AI tasks studied with this platform is robotic soccer, a competitive game very much like human, but played with autonomous mobile robots.
Robot soccer provides great opportunities for research in AI, since it raises a range of problems prototypical for many other, more serious robot applications. Annual robotic soccer competitions have attracted large numbers of AI researchers and added a lot of excitement to the field of robotics.
Human augmentation:
A final application domain of robotic technology is that of human augmentation. Researchers have developed legged walking machines that can carry people around, very much like a wheelchair. Several research efforts presently focus on the development devices that make it easier for people to walk more their arms, by providing additional forces through extra-skeletal attachments. If such devices are attached permanently, they can be thought of as artificial robotic limbs. Robotic teleoperation, or telepresence, is another form of human augmentation.
Teleoperation involves carrying out tasks over long distances, with the aid of robotic devices. A popular configuration for robotic teleoperation is the master-slave configuration, where a robot manipulator emulates the motion of a remote human operator, measured through a haptic interface. All these systems augment people’s ability to interact with their environments. Some projects go as far as replicating humans, at least at a very superficial level. Humanoid robots are now available commercially through several companies in Japan.
Object Recognition using AI:
Vision enables us to recognize people, animals, and inanimate objects reliably. In AI or computer vision, it is currently to use the term object recognition to refer to all of these abilities. This includes determining the class of particular objects that have been imaged-e.g, a face-as well as recognizing specific objects-e.g., Bill Clinton’s face. Motivating applications include the following:
Biometric identification:
Criminal investigations and access control for restricted facilities require the ability to identify unique individuals. Fingerprints, iris scans, and facial photographs result in images that must be matched to specific individuals.
Content based image retrieval:
It is easy to find a location in a document, if one exists, for the string “cat”-any text editor provides this capability. Now consider the problem of finding the subset of pixels in an image which correspond to the image of a cat. If one had this capacity, one could answer image queries such as “Bill Clinton and Nelson Mandela together”,” a skater in mid-air”, “the Eiffel Tower at night”, and so on, without having had to type in caption keywords for each photograph in a collection. As image and video collections grow, manual annotation cannot scale.
Handwriting recognition:
Vision is used to recognize not only objects, but also activities. We can identify gaits, expressions, gestures, actions and so on. Research on activity recognition is still in its infancy, so in this section we will concentrate on object recognition.
The problem of visual object recognition is generally easy for people, but has proved to be very difficult for computers. One wants to be able to identify a person’s face in spite of variations in illumination, pose with respect to the camera, and facial expression. Any of these changes causes widespread differences in pixel brightness values, so a straightforward comparison of pixels is unlikely to work. When one wants to recognize examples of a category such as “car”, one must cope with the within category variation. Even the very restricted problem of recognition of handwritten digits in postal zip codes proved to be quite a challenge.
Multiple views of three dimensional object
The challenge is image segmentation. Any image will typically contain multiple objects, so we need first to partition it into subsets of pixels that correspond to single objects. Once the image has been partitioned into regions, one can then input these regions or assemblies of regions into a classifier to determine object labels. Unfortunately, bottom up segmentation is an error prone process, so alternatively one might seek to find object groups top down. That is, search for a subset of pixels that you can classify as a face, and if you succeed, you have found a group! Purely top down approaches have high computational complexity, because one needs to examine image windows of different sizes, and at a different locations, as well as compare them to all the different object hypotheses. At present, most practical object recognition systems use such a top down strategy, though this might change as bottom-up techniques improve.
The second challenge is to ensure that the recognition process is robust against variations in illumination and pose. Humans can recognize objects in spite of considerable variation in precise appearance as measured by pixel brightness values. For examples, we can recognize a friend’s face under different illumination conditions, or at different angles of view. As an even simple example, consider recognizing the hand written digit 6. One should be able to do this at different sizes and at different positions in the image, and in spite of small rotations of the view.
The key point to note here is that geometrical transformations such as translation, scaling and rotation, or transformations of image brightness caused by moving light sources physically, have a different character than the intra-category variation such as exists between different human faces, or the different ways of writing the digit 4. On the other hand, the effects of geometric and physical transformations are systematic and one should be able to factor them out to by a proper design of the features used represent the training instances.
To provide invariance under geometric transformations, one technique that has proved quite effective is to preprocess the image region into a standard position, scale, and orientation. Alternately, we can merrily ignore the cause nature of the geometrical and physical transformations and think of them as just other sources of variability for the classifier. In the training set, example need to be provided corresponding to all these variations, and the hope is that the classifier will include an appropriate set of transformations of the input so that the variations are factored out.
The use of shape for object recognition has proved to be much more difficult. Broadly speaking, there are two main approaches: brightness based recognition, in which pixel brightness values are used directly, and brightness based recognition, which involves the use of spatial arrangements of extracted features such as edges or key points.
Brightness based recognition:
A variety of statistical methods have been used to develop face detectors from image databases, including neural networks with raw pixel inputs, decision trees with features defined by various bar and edge filters, and naïve Bayes models with wavelet features.
One negative aspect of using raw pixels as feature vectors is the great redundancy inherent in this representation. Consider two nearby pixels on the cheek of a face; they are likely to be very highly correlated because of similar geometry, illumination, etc. Data reduction techniques, such as principal component analysis, can be used successfully to reduce the dimensionality of the feature vector, enabling recognition of such things as faces with greater speed than one would get in a higher-dimensional space.
Feature based recognition:
Many different definitions have been proposed for distances between images. One of the more interesting approaches is based on the idea of deformable matching. In his classic work on growth and form, D’Arcy Thompson observed that related but not identical shapes can often be deformed into alignment using simple coordinate transformations. We operationalize a notion of shape similarity as a three stage of process:
• Solve the correspondence problem between the two shapes,
• Use the correspondence to estimate an aligning transform, and
• Compute the distance between the two shapes as a sum of matching errors between corresponding points, together with a term measuring the magnitude of the aligning transformation.
The distance between two shapes can be defined as a weighted sum of the shape context distances between corresponding points and the bending energy associated with he thin plate spline. Given this distance measure, one can use a simple nearest-neighbor classifier to solve the recognition problem.
Pose Estimation:
In addition to determining what an object is, we are also interested in determining its pose, i.e., its position and orientation with respect to the viewer. For instance, in an industrial manipulation task, the robot arm cannot pick up an object until the pose is known. In the case of rigid objects, whether three dimensional or two dimensional, this problem has a simple and well defined solution based on the alignment.
Robotics:
Introduction:
Robots are physical agents that perform tasks by manipulating the physical world. To do so, they are equipped with effectors such as legs, wheels, joints, and grippers. Effectors have a single purpose: to assert physical forces on the environment. Robots are equipped with sensors, which allow them to perceive their environment. Present day robotics employs a diverse set of sensors, including cameras and ultrasound to measure the environment, and gyroscopes and accelerometers to measure the robot’s own motion.
Most of today’s robots fall into one of three primary categories. Manipulators, or robot arms, are physically anchored to their workplace, for example in a factory assembly line or on the international space station. Manipulators are by far the most common type of industrial robots, with over a million units installed worldwide. Some mobile manipulators are used in hospitals to assist surgeons. Few car manufacturers could survive without robotic manipulators, and some manipulators have even been used generate original artwork.
The second category is the mobile robot. Mobile robots move about their environment using wheels, legs, or similar mechanisms. They have been put to use delivering food in hospitals, moving containers at loading docks, and similar tasks. Other types of mobile robots include unmanned air vehicles(UAV), commonly used for surveillance, crop-spraying, and military operations, autonomous underwater vehicles(AUV), used in deep sea exploration, and planetary rovers.
The third type is a hybrid: a mobile robot equipped with manipulators. These include the humanoid robot, whose physical design mimics the human torso. The field of robotics also includes prosthetic devices, intelligent environments, and multibody systems, wherein robotic action is achieved through swarms of small cooperating robots.
Honda´s ASIMO robot Tokyo, Japan
Real robots usually must cope with environments that are partially observable, stochastic, dynamic, and continuous. Some, but not all, robot environments are sequential and multiagent as well. Partial observability and stochastically are the result of dealing with a large, complex world. The robot cannot see around corners, and motion commands are subject to uncertainty due to gears slipping, friction, etc. Also, the real world stubbornly refuses to operate faster than real time.
Robot Hardware:
Sensors:
Sensors are the perceptual interface between robots and their environments. Passive sensors, such as cameras, are true observers of the environment: they capture signals that are generated by other sources in the environment. Active sensors, such as sonar, send energy into the environment. They rely on the fact that this energy is reflected back to the sensor. Active sensors tend to provide more information than passive sensors, but at the expense of increased power consumption and with a danger of interference when multiple active sensors are used at the same time.
Many mobile robots make use of range finders, which are sensors that measure the distance to nearby objects. One common type is the sonar sensor, also known as an ultrasonic transducer. Sonar sensors emit directional sound waves, which are reflected by objects, with some of the sound making it back into the sensor. The time and intensity of this returning signal thus carry information about the distance to nearby objects.
Effectors:
Effectors are the means by which robots move and change the shape of their bodies. To understand the design of effectors, it will help first to talk about motion and shape in the abstract, using the concept of a degree of freedom. We count one degree of freedom for each independent direction in which a robot, or one of its effectors, can move.
Sensors and effectors alone do not make a robot. A complete robot also needs a source of power to drive its effectors. The electric motor is the most popular mechanism for both manipulator actuation and locomotion, but pneumatic actuation using compressed gas and hydraulic actuation using pressurized fluids also have their application niches. Most robots also have some means of digital communication such as a wireless network. Finally, there has to be a body frame to hang all the bits and pieces on and a soldering iron for when the bits and pieces stop working.
Methods of AI:
Perfect rationality:
Perfectly rational agents acts at every instant in such a way as to maximize its expected utility, giving the information it has acquired from the environment. We have seen that the calculations necessary to achieve perfect rationally in most environment are too time-consuming, so perfect rationally is not a realistic goal.
Calculative rationality:
This is the notion of rationally that we have used implicitly in designing logical and decision-theoretic agents. A calculative rational agent eventually returns what would been the rational choice at the beginning of its deliberation. This is an interesting properly for a system to exhibit, but in most environments, the right answer at the wrong time is of no value. In practice, AI system designers are forced to compromise on decision quality to obtain reasonable overall performance; unfortunately, the theoretical basis of calculative rationality does not provide a well-founded way to make such compromises.
Bounded rationality:
Herbert Simon (1957) rejected the notion of perfect rationality and replaced it with bounded rationality, a descriptive theory of decision making by real agents.
He suggested that bounded rationally works primarily by satisficing that is, deliberating only long enough to come up with an answer that is “good enough”. Simon won the Nobel Prize in economics for this work and has written about it in depth. It appears to be a useful model of human behaviors in many cases. It is not a formal specification for intelligent agents, however, because the definition of “good enough” is not given by the theory. Furthermore, satisficing seems to be just one of a large range of methods used to cope with bounded resources.
Bounded optimality:
A bounded optimal agent behaves as well as possible, given its computational resources. That is, the expected utility of the agent program for a bounded optimal agent is at least as high as the expected utility of any other agent program running on the same machine.
Of these four possibilities, bounded optimality seems to offers the best hope for a strong theoretical foundation for AI. It has the advantage of being possible to achieve: there is always at least one best program-something that perfect rationality lacks. Bounded optimal agents are actually useful in the real world, whereas calculatively rational agents usually are not, and satisficing agents might or might not be, depending on their own whims.
The traditional approach in AI has been to start with calculative rationality and then make comprises to meet resource constraints. If the problems imposed by the constraints are minor, one would expect the final design to be similar to a BO agent design. But as the resource constraints become more critical-e.g., as the environment becomes more complex-one would expect the two designs to diverge. In the theory of bounded optimality, these constraints can be handled in a principle fashion.
In sum, the concept of bounded optimality is proposed as a formal task for AI research that is both well defined and feasible. Bounded optimality specifies optimal programs rather than optimal actions. Actions are, after all, generated by programs, and it is over programs that designers have control.
Success of AI:
Intelligent computers more powerful, but will that power be used for good or ill? Those who strive to develop AI have a responsibility to see that the impact of their work is a positive one. The scope of the impact will depend on the degree of success of AI. Even modest successes in AI have already changed the ways in which computer science is taught and software development is practiced. AI has made possible new applications such as speech recognition systems, inventory control systems, surveillance systems, robots, and search engines.
Finally, it seems likely that a large-scale success in AI-the creation of human-level intelligence and beyond-would change the lives of a majority of humankind. The very nature of our work and play would be altered, as would our view of intelligence, consciousness, and the future destiny of the human race. At this level, AI systems could pose a more direct threat to human autonomy, freedom and even survival. For these reasons, we cannot divorce AI research from its ethical consequences.
Which way will the future go? Science fiction authors seem to favor dystopian futures over utopian ones, probably because they make for more interesting plots. But so far, AI seems to fit in with other revolutionary technologies whose negative repercussions are out weighted by their positive aspects.
No comments:
Post a Comment