New style of artificial intelligence: behavioral AI research
Li Jianhui
I. Classic artificial intelligence research and the problem encountered
Can the machine think? 1950 famous computer expert Alan Turning published an epochial paper in "Mind) magazine:" Computing Machine and Intelligence ". In this paper, Tuling believes that the problem that the machine can think about should be used instead of whether the machine can be replaced by the famous "Tuling Test" of his design. If the machine can pass this test, it can be said that the machine has thinking.
The rich imagination and enrollment in this article have attracted people's intelligence of machine intelligence. In 1956, ten mathematician and logologians held a summer academic seminar in Dartmouth College in New Hampshire, USA. Their goal is to determine the characteristics of intelligence to simulate. In the participants, John McCarthy, Mawen Minsky, Herbert Simon, and Allen Newell. These four people have laid the cornerstone of artificial intelligence.
Minsky defines artificial intelligence as "Let the machine's intelligence of the need for people's intelligence". Some people think that artificial intelligence is to build a useful intelligent system and understand people's intelligence. In the early days of the birth of the computer, it is mainly used to help people perform numerical calculations. In this regard, the computer exhibits excellent ability. At the same time, people have also begun to try smart behaviors such as chess, translation language and theorem certification. In this regard, the computer also performs extraordinary. In 1976, Apel, University of Illinos, used 1200 hours in three computers, made 20 billion logic judgments, and finally proved that people have always wanted to prove and have not proved four-color theorem. This result has caused a sensation in the world. It shows that in a special field, the computer's computing capacity far exceeds humans. In this context, the research of artificial intelligence after the 1970s has two hot spots of expert systems and knowledge engineering from theoretical steering.
The 70s expert system developed, many different expert systems in a series of different fields, such as Simon's Students Edward? Fergen Bum and Nobel Prize winners, Genetologists Schubaya Leadeberg cooperation The development of the Endral system that can be determined to determine the unknown organic compound structure, Stanford University Research Medical Advisory System, MYCIN, Pittsburgh University, developed disease diagnostic system INTERNIST, etc. They use specialized knowledge bases, interrelationships, and specialized reasoning rules to complete their respective work.
Since the knowledge of the expert system is input by the person, it is entered into computer, so the workload is very huge. In this way, the acquisition of knowledge becomes a bottleneck of the development expert system. The computer is still far from the same ability. Although the high level of intelligence has approximate expert's level, it is not as good as a 3-year-old child, such as the recognition of the number of people, and the recognition of visual information. Common sense reason, uncertain knowledge and intuition thinking. In this context, the knowledge engineering came into being. Knowledge project is to study the structure, classification, prediction, access, acquisition, transmission, conversion, management, utilization, proliferation, learning and expression of human knowledge, classify, predict, access, acquire, transmit, transform, manage, utilize, proliferate, learn, and expire.
However, in the 1980s, artificial intelligence and robotic manufacturing have encountered unprecedented difficulties. People find that expressing the knowledge of human experts in the form of a simple program is much more difficult than imagination. Therefore, the expert system is not as successful as people initially imagined. Of course, this is not to say that the expert system has stopped working, and many expert systems are still in use, but less than expected. And these experts in these uses are often in some less complex areas.
The progress of robotic manufacturing related to artificial intelligence is not as fast as the initial imagination. In 1966, Minski and his colleagues in MIT believe that they can figure out the work mechanisms of visual work within one year. However, they did not succeed. As a result, Minsky is aware of the complexity of the problem. To simplify this problem, they turn the test design to build a house or build a tower of the robot. What is expected to be in Mensky is that this "microcomputer world" research is also unsuccessful. The robot's vision system is not good in identifying the toy module; the motor control program cannot set a toy module accurately; the robot is very awkward when the robot is placed. However, their work is not a little bit. Their work shows that in a simple world, the robot can translate what it see, can control the module, and even answer some questions about these modules. MIT's researchers once extended their things from the "Module World" to the real world, but due to the overall progress of robotics, the funding of research funds began to decrease, and the program has not been possible. These people realize that intelligence is far from being as simple as they initially imagined. "In contrast to the expectations, micro-world research methods do not guide gradual solution to general intelligence issues (CREVIER 1993: 114)." Faced with these difficulties, some experts believe that this is mainly related to computer. They believe that with more powerful computers, computer intelligence will reach and exceed humans.
Other experts believe that the difficulties encountered by artificial intelligence and robotics have a fundamental flaw in traditional artificial intelligence and robotics research. To make artificial intelligence and robotics have new development, there must be a new research idea. Just at this time, the study of artificial life is rising. Artificial life provides a new way of thinking for robotics.
Second, the rise of behavioral AI
Traditional artificial intelligence research and robot manufacturing are from top-down research methods, that is, researchers first identify a complex high-level cognitive task, then decompose this task into a series of subtasks, and then construct the integrity of these tasks. system. This study is stored in advance, then uses the computer's large-capacity storage capacity and rapid calculation capabilities to process relevant knowledge. Therefore, this study is called "knowledge-based research". With the speed, capabilities and various hardware capabilities of the computer, traditional artificial intelligence has achieved some success in some special fields, such as "Deep Blue", which has been able to defeat the chess master. But "dark blue" chess is very different from human chess. Human Chess Masters often rely on intuition to move their chess, and "dark blue" relies on its fast computing power, searching and picking out from thousands of possible ways. Therefore, it is not the same as the intelligent intelligent intelligent intelligent intelligence of artificial intelligence systems based on knowledge. Human intelligence, and even more intelligent demonstrated in the process of external environmental response is more flexible and natural than intelligence than existing robots. Therefore, some robot experts have begun to consider the biological learning to nature, and how the organics in nature are well completed, and there is no suitable work.
From the perspective of biological evolution, people's intelligence is not sudden, but experience a series of intermediate development phases. This development can be seen from many animals. Therefore, some robot experts believe that the awareness of these lower-level biological intelligence may help us understand how people with higher levels of human beings are organized. Therefore, in contrast to traditional artificial intelligence research, new methods take self-disciplined research strategies to focus on physical systems that can perform various tasks in real world, such as mobile robots.
Wilson (S. Wilson) can be said to be one of the researchers entering this field. Wilson believes that intelligence is closely related to the living requirements of survival. It is the driving force for survival to continuously define different problems in nature, thereby producing diversity. Therefore, Wilson believes that only the intelligence starting from replica can finally replicate people's intelligence. As a result, Wilson's role in manufacturing can avoid danger when conducting robotics research, looking for food, can use animals to deal with robots. Wilson named his artificial creatures "Animat". Wilson hosted a number of international seminars on Animat, so animat was quickly opened in artificial intelligence and artificial life. Brooks (Rodney Brooks) also begins with new vision from the beginning of the 1980s to view scholars in artificial intelligence and robotics. Brooks found in the study, traditional logic programs appeared very slow and clumsy in the navigation of robots, so he intends to find a new method for building robots. Brooks believes that traditional robot research first assumes that the real world is a still world, when we solve all the problems in this still world, then return to the dynamic world to study it. But Brooks assumes that the world is dynamic, so it can avoid falling into an endless operation. Brox's goal is to manufacture mobile robots that can handle changes in daily transactions, so he is concerned about the behavior of robots. Brooks believes that in order to truly inspect the concept of intelligence, it is important to construct a complete robot capable of interacting in dynamic environments and outside. The internal model of the external world is fully characterized, that is, the traditional artificial intelligence model is difficult to establish, on the other hand, the behavior of the robot is unnecessary.
Based on this understanding, Brooks believes that new research should emphasize the following aspects (Brooks 1991): (1). The robot is placed in the real world that does not deal with the abstract description, but directly reacts its exterior world. (2) Subcutaneized. The robot has the body and uses the body to experience the dynamic external world. (3) Intelligent. These robots look into intelligence, but intelligence is not just from its computing engine, and also from the environmental world, the signal transition in the inductor, and the interaction of the robot and environment. (4) Exemption. The dynamic interaction between the robot and its environment and the dynamic interaction between the parts of the robot can be surprisingly structural and functional. Intelligence is a result of being out. This new artificial intelligence research idea, the rise in the middle and late 1980s, some people call it "Autonomous Agent Research", and some people call "behavioral AI" or "self-bottom AI" To distinguish between classic "knowledge-based AI", or "from top-down AI". That is, at this time, Randon launched an international seminar of the first artificial life, announced a new discipline: the birth of artificial life. Randon believes that the ultimate goal of artificial life is to establish intelligent artificial life. But artificial life is not directly going to intelligence, but starts with the study of low-rise life. The presence of some low-level life components prompts us to find new methods to solve intelligence issues, thereby planning to logically replace other methods. It can be seen that Randon's feelings and Wilson, Brooks et al., Through the study of relatively simple biological behavior to make robots with low-level intelligence ideas. Because of this, behavior-based artificial intelligence research is integrated with artificial life, and has become an important part of artificial life. They replenish each other, mutual promotion: On the one hand, artificial life provides a new theoretical basis and vast activity space based on behavioral AI in scientific legality; on the other hand, behavioral-based AI also enriches artificial life. Studies have become an important part of artificial life. Third, the accommodation structure of independent robots
Since traditional artificial intelligence is concentrated in "reasoning",, when building robots, the "Sense-Model-Plan-Act) frame is taken when building robots. As mentioned earlier, early robots were placed in a simple artificial world, they feel this world, then think about it, strive to build two dimensional or three-dimensional models of this world. Then, they make plans based on these models and make robots to achieve specific goals by this plan.
But Brooz believes that the robot does not have to be more complicated from feeling, only two steps can: ie feel, then action. The robot feels something, then responds to this feeling, there is no need to have "build models" and "make a plan". How to get this kind of thinking to reality? Brooks regards the different behaviors of the robot as a similar modular thing that is closely intertwined. The robot selects the appropriate behavior in a given time according to its sensory acceptance. Essentially, robot behavior is similar to a huge limited automat machine. Information about the environment and current state of the robot will be processed according to rules, and these rules run in parallel, and robots have an operation from these operations that may result in a series of continuous activities.
Brooz believes that using these new ideas, new robots can abandon the complex planning, mapping and cognition required by traditional artificial intelligence paradigms. The new robot has many modules layers that can trigger their behavior at appropriate time. The top behavior module of the module layer may be a "explore" module; "walking" module below it; then it may be a group of behaviors that are determined by the leg sensor input by the leg sensor. Unlike traditional from top, Brooks' robots start upwards from the lowest layer. Since these processes allow a behavioral tolerance to control another behavior, for example, using low-level behavior, let the robots pay the external world, use high-level behaviors to seek targets, Brooks, said his theoretical framework called "inclusive structure" . The intellier structure is completely different from the central control mode of traditional artificial intelligence, with the following features: (1) There is no location about the central model of the external world; (2) The robot's perception, central processing and brake system are closely linked; (3) By adding new more behavior networks in existing networks, it can enhance the ability of the system; (3) There is no high and low levels between hierarchies; (4) Various behaviors work in parallel. Brooks believes that this inclusive structure avoids the cognitive bottleneck of traditional artificial intelligence research framework, and it can be built to build a complicated structure. He believes that using this inclusive structure, there will be no obstacles to prevent us from building more and smarter robots, including human level intelligence.
Four, Allen, Herbert and Renges
Brooks and the researchers in his laboratory use this inclusive structure to design multiple robots, the earliest is "Allen". However, Allen's initial success is mainly by computer simulation rather than manufacturing by actual robots. Allen has three behavior layers. The first layer is used to avoid obstacles, the second layer is used for random swing, and the third layer is used to move toward distant distances. Allen can walk along the wall, you can identify the door, but because it is running on the LISP machine, Allen does not become a completely independent robot.
The second robot manufactured by Brooks's inclusive structure is called "Herbert" (named according to the name of the artificial intelligence Herbert Simon). Herbert's inclusive structure operates from a 24-pixed 8-bit processor. Herbert is equipped with 30 infrared sensors that help it avoid obstacles. It relies on a laser-based visual system to identify objects. It also has a robot for gripping things. Herbert can walk from the information obtained from the sensor, walk along the wall, can avoid obstacles, or pick up the soda on the ground.
The best reflecting Brooks robot manufacturing idea is "Genghis". Renges is a Six legs "robot" manufactured by Brooks laboratory researchers. Its body is a metal chassis with a computer chip; its legs are a metal rod with rubber sleeves; there is a row of six like lights on the head; two hard images from its chest Bouvet wire. The reason why Renges is so named because it can be stepped on other objects. Although root's six legs can be coordinated to move forward, but each leg is independently controlled. Each leg is given several simple behavioral rules so that they know how to move in different situations.
After the basic walking behavior is created, Brooks et al. Start again to add more sensors and behavior layers for roen. Some of these inductors monitors the angle of the body tilt when root sway, and some other sensors track the size of the legs in the swing. Renges uses information from these inductors, producing some new behaviors, these new behaviors can help it go better. When adding new behaviors, the researchers do not modify the original behavior. New behaviors are only covered in basic behaviors when conditional needs. These high-level behaviors will not work without need. This inclusive structure of Renges has greatly enhanced its capacity. Now Rigs can not only walk on a flat ground, but also walk on a high and low ground, even walking on the slope, but also obstacles like telephone numbers. In this process, no one "tells" How to move when you encounter special circumstances such as telephone number, ramp, but it can find a solution to the problem when encountering these situations. These seemingly intelligent activities in Renges are actually exacerbated from various activities without intelligence. Therefore, root behavior has distinguished Brooes an important design idea: complex intelligent behavior is "out" from a simple rule. Rigs has been MIT mobile robot laboratory in 1990. Later, roots were replaced by the machine animal - Attera, which was more active.
Five, Test Ge: Based on behavioral humanoid robot
Brooks put forward his new insights in criticism of traditional artificial intelligence research, so, although he made a remarkable achievement in machine insects (he also called Robot), he was also Some artificial intelligent scholars who persist in traditional research methods. These people have said that the work made by Broo is simplicity, and the intelligence of a low simulation is easy, but this simulation is not related to people's intelligence. Brooks and their followers set a lower standard for themselves, so they achieved some no disappointing results and is not amazing. The meaning of these people is to say that traditional artificial intelligence studies have not further achieved remarkable achievements, just because researchers are too high for their own goals. In the face of this challenge, Brooks began to think about the use of his approach to make a class robot. His latest research project "CoG" (COG) is to meet this challenge.
Similar to the machine insects made from Brooks and their graduate students, the intelligence of the exam will also be intelligent in direct interactions with the outside world. Different, the exam is a robot with a similar person's appearance. With Brooks, a hypothesis behind the creation of robotic exams with people is: Classic intelligence requires robots to interact with the outside world. Specifically, Brooks believes that there are two reasons for manufacturing robots with a class. First, our body form is critical to characterizing our inner thoughts and our language. If we want to make a robot with people, it needs to have a similar person's body to develop similar species. However, because the body of the robot we can only can only be a rough approximation of a human body, it must pay attention to grasp the essence of the human body, and cannot just simulate the irrelevant aspect of the human body. Second, a robot with a human appearance can more naturally interact with such robots more naturally. Because an important aspect of humans is interacting with other humans. If the robot has a human look, then the interaction with it will be easier and more natural. Brox and others have observed that as long as the class robots issued a little similar person's signal, people will naturally enter the interaction mode of people. From this interaction, we can get many cases of dynamic interactions with robots. These examples are critical to accumulating the experience of manufacturing a robot with a human smart.
To date, the exam is equipped with torso, arms, necks, and heads with eyes and ears. The torso is placed on a stand that allows it to rotate as much as possible. Most of the body is equipped with gears and follow-up systems to the arm. Body and arms are covered with rubber skin. The rubber skin is equipped with a sensor, so that the exam knows what it is doing. The neck of the exam can turn it, which makes it see what it is interested. Its two eyes can be coordinated, or it can be telescopically and zoom. Its head uses a vestibular system that is similar to human beings that can be balanced automatically. The exam also has a brain, but unlike the single sequence processor used by traditional artificial intelligence researchers, Test Qigu uses multiple parallel processors to make the cerebral brain more natural. The human brain works like a parallel computer because it can handle many different sensory information at the same time. The parallel processor of the Codge Brain will help it coordinate the activities of each independent component. The coordination of necks, heads and eyes requires it to have quite parallel processing capabilities. Different feelings of Digger also require different processing systems. This is parallel to our human perception system. Our different feelings are closely linked. Various feelings, such as smell, tactile, visual, sound, taste, etc., together, form awareness of one thing. The cognitive system of the exam started from the feeling. Brooks said, "We start building a class robot from this sensory level, all intelligence will be based on the calculation of the information derived from the information or from the feeling (Brooks & Stein 1993: 4)." The study is still in progress, it has no legs, not free to move. Perhaps it is aware of the difficulties of the replica human smart, Brooz said that he doesn't plan to make the exam to have an adult level of intelligence. If the exam is able to achieve the smart similar to two-year-old children. Is Brooke reduced his own standard? Can Brooks succeed? Some artificial intelligent builders who have a central control paradigm hopes to see Brooks fail, because this can explain that Brooks' research methods have fundamental defects. Minski, the founder of the artificial intelligence laboratory where Brooks, Brooz refused his robot to combine the traditional artificial intelligence program to handle abstract categories such as time or physical entity, which undoubtedly make his robot There is no use value. It is also believed that Brooks' research is too trivial. "Returning to the 1950s, reflective behavior can help robots avoid hitting the wall, but we need a high-level intelligence now, which makes robots to turn left or right when they encounter crossroads. Determined intelligence (turn from Friedman, 2001 Chinese version: 20). "
Brox himself thinks that only time can tell people which method is correct. He said, "I am correct or wrong will be an empirical problem (swap from Ward 1999: 167)." If Brooks is further successful, then it will prove the Brooks method; if Brooke failed Then, in the experience, Brooks' methods are defective. So, make the final judgment, we also have to wait. However, from the current point of view, Brooks has achieved achievements in robotics exceeding the achievements of the traditional AI mode. Now he is the director of the MIT Artificial Intelligence Lab, not just the director of the Mobile Robot Lab, which reflects artificial intelligence research from artificial intelligence from one aspect, has become one of the mainstream schools of artificial intelligence.
references
1. Brooks, Rodney (1991). "" MIT AI MEMO NO. 1293.2. Brooks, Rodney (2001). "The Relationship Between Matter and Life." Nature, 409: 409-411.3. Brooks, Rodney, & Stein, Lynn Andrea (1993). "Mit Ai Memo No. 1439.4. CREVIER, DANIEL (1993). AI: The Tumultuous History of The Search for Artificial Intelligence. New York: Basic Books.5. Langton, CG (1989). "Artificial Life", From Artificial Life. Chris Langton, Ed. SFI Studies in The Sciences of Complexity, Proc. Vol. Vi. 1989. Redwood City, CA: Addison-Wesley. Reprinted in Boden 1996). Levy, Steven (1992). Artificial Life: a Report from the Frontier Where Computers Meet Biology. New York: Vintage Books.7. The cog shop Website: http://www.ai.mit.edu/ Projects / cog.8. The Mit Ai Lab Mobot Group: http://www.ai.mit.edu/projects/mobile-robots/ Robots.html.9. Turing, AM (1950). "Computing Machinery and Intelligence. "Mind, 59: 433-6010. Ward, Mark (1999). Virtual Organism: The StartL Ing Word of Artificial Life. New York: Thomas Dunne Books.11. Friedman (2001), "Make the brain: creating a human brain and energetic intelligence". Zhang Mo and other translations. Beijing: Sanlian Bookstore. This article originally filed Hong Kong "21st Century" Aug. 2003. After adding content, it is blended in the book "Digital-Digital Age Artificial Creating Life" (China Bookster Publishing House 2004).
