(World Chess Champion Kasparov in the fourth game against the IBM Deep Blue Computer. Figure: Stan Honda Getty Images)
Products | NetEase Scientific People Section
Translator | Sun Wenwen
Twenty years ago, the "dark blue" computer developed by IBM became the first machine to beat the chess champion and was known all over the world. The victory of computers over humanity has caused people to worry about their being surpassed by machines. Of course, Kasparov and others have also been skeptical that IBM has won through cheating. However, what happened in May 1997 was far more revolutionary than the revolution. Behind it was a story about the arduous progress of smart machines and the rise of brave people who lost first.
For decades, computer scientists have considered chess as an instrument bar for testing artificial intelligence. Computers that specialize in chess emerged in the late 1970s, but it took ten years until a team of students at Carnegie Mellon University built a computer called Deep Thought and at Beat the human master in the regular season. The victory did not last long, and later in the same year, chess player Kasparov easily defeated “thinking†in two games. The technology of the Cameron team left a deep impression on IBM. The Blue Giant recruited team members to develop an early version of Deep Blue. It can be seen from the name that "dark blue" is both the evolution of "thinking" and the color of IBM itself. After a preliminary evolution of the "deep blue" in 1996 Philadelphia again challenged Kasparov. Unfortunately, the artificial intelligence in the second matchdown defeated again, winning only one of six games.
Murray Campbell, the AI ​​expert behind Deep Blue, described the seemingly minor victory as “very important to us and shows that we are on the right track.†Campbell is now a well-known researcher at the cognitive computing facility at the IBM TJ Watson Research Center.
"Before the final confrontation in 1997, we have made enough improvements to the system based on our past experience." Campbell received an interview with "Scientific American" about the long-term obsession of computer scientists' objects, how IBM has captured chess and AI. Future challenges.
(The following is edited)
How did you first participate in the "dark blue" project?
I am a member of the graduate team at Carnegie Mellon University that IBM has compiled. I was originally interested in computer chess and wrote a chess program during my undergraduate study. My job at the university is a very broad artificial intelligence research, not a chess computer designed to beat the world champion. However, as an amateur project, many people on the team (including Xu Xiongfeng and Thomas Anantharaman) set out to develop a machine called “Thinking†and successfully defeated a chess master in a professional-level competition.
Later, IBM noticed our success in building a computer with a low budget, and let a group of us join IBM Research (at the end of 1989) to develop the next version of “Thinkingâ€, which is the later “dark blueâ€. They want to know what superiority of chess players is in the foreseeable future. Although other researchers believe it will take decades to complete, our feeling is that it will take only a few years.
How does chess attract computer scientists?
The number of people playing chess in the world is in the hundreds of millions. This kind of game requires strategy, foresight and logic. These are the elements that make up human intelligence. So it makes sense to use chess to measure the level of development of artificial intelligence.
For a chess game like this, we would say "yes, the computer can certainly do well because it is a well-defined game - rules, moves and goals." In a limited question, you know all the information. However, despite this simplification, chess is still a very complicated game, which is why the computer field has used a full 50 years of development to beat the world champion.
How is your role in the dark blue team?
I am an AI expert. The AI ​​in 1989 and early 1990 was very different from today. Due to the lack of machine learning, it was mainly what we now call "old-style AI" or "symbolic artificial intelligence." Although machine learning was a serious area at the time, it was not like today's huge datasets and large computers and very advanced algorithms to understand the data, and to extract amazing operational algorithm models. When I joined IBM at that time, AI was quite primitive. The algorithms we use to search the database also have logic, planning, and reasoning. These are all necessary for deep blue competition.
What were the main limitations of AI at the time?
The hardware at the time was not sufficient to support large networks that handle large amounts of data, and the data did not meet the demand. Whenever you look back at the most popular computer systems 20 or 25 years ago, the way these systems work will surprise you. However, we still did it.
At that time, people did not realize the benefits of data. It is useless to build a very large data set, not to mention the processing power at that time does not have to use so much data. So we use a smaller data set.
How can your own knowledge of chess help in creating a "deep blue" process?
It may not be as useful as you think. In the early stages, I could also identify system errors and know how to come up with solutions to avoid extraneous growth. Although this is enough to help us achieve a certain level, the real game requires real professional chess knowledge. So later we invited a chess expert such as Joel Benjamin to help.
How can chess masters help "dark blue" growth?
Their help can be divided into two aspects. One is to help build the opening database, which helps the chess program save time and make a reasonable move. For hundreds of years, humans have studied the beginnings of various chess games and developed their own favorite moves. The masters helped us choose some of them to be programmed into "dark blue".
On the other hand, the masters are also partners in deep blue practice. The master plays chess with the computer and tries to find out the flaws of the system. The master then discusses with the program developer to find out where the defect is and try to solve it. Sometimes, taking into account our limitations - we have programmed some of our computer instructions directly onto a chip called a chess accelerator instead of software - there are some problems that cannot be solved easily. But usually we have ways to improve the machine's ability to handle known problems.
How does Deep Blue decide how to go each step?
Dark Blue is a hybrid decision. It combines a universal supercomputer processor with a chess accelerator chip. The software running on the supercomputer performs a portion of the calculations, more complex moves to the accelerator, and then calculates possible moves and results. The supercomputer determines the final move based on these results.
From the failure of 1996 to the success of revenge in 1997, during the "dark blue" made what improvements?
We did a few things. By creating a new generation of hardware, we have almost double the speed of operation. Then, we improved the chess knowledge of the system by improving the chess chip, so that it can identify different chess games, and the object chess concept has a better understanding. These chips find the best moves for different situations from a multitude of possibilities. One of the improvements during the period was that we found more patterns in a chess game, assigned them, and thus made a more accurate assessment of the situation. The 1997 version of Deep Blue retrieved 100 million to 200 million chess games per second, depending on the type of game. In some cases, the system can calculate 6 to 8 moves, up to 20 or more. However, although we are confident that the "Deep Blue" in 1997 is much better than the 1996 version, I think the most likely result of the game is a draw. Even in the final game of the game, I am still looking forward to a draw and then go to another day.
Why did IBM reject Kasparov's request to fight after the game?
We feel that our goal of demonstrating that computers are sufficient to defeat the world championship has been achieved and it is time to enter other important research areas.
What have been the changes in AI since the competition has been 20 years ago?
Of course, the machine processing speed and memory have improved. People began to collect more data, which provided training data for the current machine learning algorithm. In the end we realized that combining these things can produce some great results. The IBM "Watson" system, which participated in the TV competition program, used machine learning to absorb the world's massive data - including Wikipedia - and used the data to answer questions about the real world. After that, we have also begun to learn how to do some perceptual tasks such as speech recognition and machine vision. This allowed Watson to perform more business-related tasks such as analyzing radiology images and sharing information with physicians.
How has the experience of participating in "dark blue" affected your subsequent work on AI?
One of the many lessons we have learned is to look at complex issues in more than one way. For example, in chess, the human way is based on pattern recognition and intuition, while the machine uses the intensive search to retrieve millions or even billions of possibilities. Usually, these methods are complementary to each other. Not only in chess, but in the face of many problems in the real world, the combination of computers and human brains must outweigh the paranoid side. For example, we do not expect computers to take full responsibility for the diagnosis and treatment of patients because the diagnosis also involves a large amount of information that cannot be digitized. However, the computer system may be very valuable in providing medical advice.
An important work we are currently doing is trying to make the neural network system capable of interpreting itself. These systems are very advanced, but at the same time they are "black boxes," and humans cannot understand how machines make such wise decisions. If the machine can't explain itself, how can people really believe it? These black box neural network systems are extremely complex and involve millions of parameters. One way to overcome this complexity may be to add training to interpret cases in the neural network training process. In computer-assisted medicine, if the system proposes a convincing reason for recommending a treatment plan, we may be able to confidently involve the machine in further medical decision-making.
Portable Power Station also known as Portable Power Bank. Our product regulated DC Power, AC Pass-Through Charging, Wireless Fast Charger, Bluetooth Playback, Attentive Start, Car charger, Type-C ....... Large capacity meet the needs of all kinds of equipment.
Portable Power Station,Indoor Portable Power Station,Solar Portable Power Station,Portable Power Station With Bms
Jiangsu Zhitai New Energy Technology Co.,Ltd , https://www.zttall.com