AlphaGo facts for kids

Quick facts for kids
AlphaGo

Developer(s)	Google DeepMind
Type	Computer Go software

AlphaGo is a special computer program that plays the board game called Go. It was created by a company in London called DeepMind Technologies, which is part of Google.

Over time, different versions of AlphaGo became even stronger. One version, known as Master, competed against top human players. After its competitive career, AlphaGo Master was followed by an even more powerful version called AlphaGo Zero. This version learned completely by itself, without studying any human games. AlphaGo Zero then led to a program called AlphaZero, which could play other games like chess and shogi. Later, a program called MuZero was developed, which learns without even being told the rules of a game.

AlphaGo and its later versions use a special method called Monte Carlo tree search. This helps them choose their moves based on what they've learned through machine learning. They use something called an artificial neural network (a type of deep learning) which is trained by playing many games, both against humans and other computers. This neural network learns to pick the best moves and predict how likely it is to win. As the network gets better, the program chooses stronger moves.

In October 2015, the first AlphaGo program made history. It beat Fan Hui, a professional Go player, without any special advantages (called a handicap). This was the first time a computer Go program had ever done this on a full-sized 19x19 Go board.

Then, in March 2016, AlphaGo played against Lee Sedol in a five-game match. Lee Sedol was one of the best Go players in the world, a "9-dan" professional. AlphaGo won the match 4 games to 1. Even though Lee Sedol won one game, AlphaGo won the series. Because of this amazing win, AlphaGo was given an honorary 9-dan rank by the Korea Baduk Association. A documentary film called AlphaGo was made about this match. The win was even chosen as one of the "Breakthroughs of the Year" by the journal Science in 2016.

At the 2017 Future of Go Summit, the Master version of AlphaGo beat Ke Jie, who was the world's number one Go player at the time. After this, AlphaGo was given a professional 9-dan rank by the Chinese Weiqi Association.

After beating Ke Jie, DeepMind decided to retire AlphaGo from competitive play. They wanted to focus on other areas of AI research. The self-taught AlphaGo Zero later beat the earlier AlphaGo version 100-0. Its successor, AlphaZero, was seen as the world's best Go player by the end of the 2010s.

How AlphaGo Was Created

Go is a very complex game for computers to play well. It's much harder than games like chess. This is because Go has many more possible moves at each turn, making it difficult for computers to plan ahead using traditional methods. Also, it's hard to teach a computer how to "evaluate" or understand the value of different positions in Go.

About 20 years after IBM's computer Deep Blue beat the world chess champion in 1997, the best Go programs were still not as good as professional human players without a handicap. In 2012, a program called Zen beat a professional player with a handicap. In 2013, Crazy Stone also beat a professional with a handicap.

Around 2014, DeepMind started the AlphaGo project. They wanted to see how well a deep learning program could play Go. AlphaGo was a huge step forward. In 500 games against other strong Go programs like Crazy Stone and Zen, AlphaGo (running on one computer) won almost every game. When AlphaGo ran on many computers together, it won all 500 games against other programs. This powerful version used 1,202 CPUs and 176 GPUs.

AlphaGo Beats Fan Hui

In October 2015, the powerful version of AlphaGo played against Fan Hui, who was the European Go champion. Fan Hui was a 2-dan professional player. AlphaGo won all five games, 5-0. This was a historic moment because it was the first time a computer had beaten a professional human Go player on a full-sized board without any handicap. The news was kept secret until January 2016, when a scientific paper about AlphaGo's methods was published.

AlphaGo Versus Lee Sedol

AlphaGo then played against Lee Sedol, a 9-dan professional from South Korea. Many people considered him one of the best Go players in the world. They played five games in Seoul, South Korea, in March 2016. The games were shown live online.

AlphaGo won four out of the five games. Lee Sedol won the fourth game, making him the only human player to beat AlphaGo in its 74 official games. AlphaGo used Google's cloud computing power, with servers in the United States. The rules used were Chinese rules, and each player had two hours to think.

AlphaGo won the first three games when Lee resigned. However, Lee surprised everyone by winning the fourth game. But AlphaGo came back to win the fifth game, making the final score 4-1.

The prize for winning was US$1 million. Since AlphaGo won, the money was given to charities, including UNICEF. Lee Sedol received $150,000 for playing and an extra $20,000 for his win in Game 4.

Later, one of the DeepMind team members, Aja Huang, explained why AlphaGo lost the fourth game. Lee Sedol's move 78, which many called the "divine move," confused AlphaGo. The program's way of finding the best moves didn't guide it correctly after that unexpected move.

AlphaGo Master's Online Games

In late 2016, a new online player appeared on Go servers, first named "Magister" and then "Master." This player quickly won 60 games in a row against many top professional players, including Ke Jie, the world's number one. DeepMind later confirmed that "Master" was an updated version of AlphaGo, called AlphaGo Master.

Go experts were amazed by AlphaGo Master's performance and its unique playing style. Ke Jie said, "After humanity spent thousands of years improving our tactics, computers tell us that humans are completely wrong... I would go as far as to say not a single human has touched the edge of the truth of Go."

The Future of Go Summit

In May 2017, the AlphaGo Master version played three games against Ke Jie, the world's top-ranked player, at the Future of Go Summit in China. Google DeepMind offered a $1.5 million prize for the winner. AlphaGo Master won all three games against Ke Jie. After this, the Chinese Weiqi Association also gave AlphaGo a professional 9-dan rank.

After winning against Ke Jie, AlphaGo officially retired from competitive play. DeepMind then stopped the team working on AlphaGo to focus on other areas of AI research. As a gift to the Go community, DeepMind released 50 full games played between different AlphaGo versions.

AlphaGo Zero and AlphaZero

On October 19, 2017, AlphaGo's team announced AlphaGo Zero. This new version was even stronger and learned without any human data. It learned by playing games against itself. AlphaGo Zero became stronger than the version that beat Lee Sedol in just three days. It reached the level of AlphaGo Master in 21 days and became better than all older versions in 40 days.

On December 5, 2017, DeepMind announced AlphaZero. This program took AlphaGo Zero's learning method and applied it to other games. Within 24 hours, AlphaZero became superhuman at chess, shogi, and Go. It beat the world's best computer programs for each of these games.

AlphaGo as a Teaching Tool

In December 2017, DeepMind released an AlphaGo teaching tool. This tool helps Go players understand the best opening moves and how likely they are to win. It uses AlphaGo Master's calculations from 10 million game simulations. This helps human players learn new strategies.

How AlphaGo Works

AlphaGo's technology combines machine learning and tree search. It learns from many games, both human and computer. It uses a method called Monte Carlo tree search, guided by two special parts called a "value network" and a "policy network." Both of these are built using deep neural network technology.

The program's neural networks first learned by watching how human experts played. AlphaGo studied about 30 million moves from past games to copy human play. Once it became good at this, it learned even more by playing countless games against itself. This process is called reinforcement learning. To be polite, AlphaGo was programmed to resign (give up) if it thought its chance of winning dropped below 20%.

AlphaGo's Playing Style

Toby Manning, who was a referee for AlphaGo's match against Fan Hui, said AlphaGo's style was "conservative." AlphaGo always tried to maximize its chance of winning, even if it meant winning by a small margin. This is different from human players, who often try to gain as much territory as possible. This explains some of AlphaGo's moves that looked strange at first but made sense later. It often made opening moves that humans rarely or never used. It also liked to use "shoulder hits," especially when an opponent was too focused in one area.

Impacts of AlphaGo's Victories

On the AI Community

AlphaGo's victory in March 2016 was a huge step for artificial intelligence research. Experts thought it would take at least five to ten more years for a computer to beat a Go champion. Most people expected Lee Sedol to win against AlphaGo.

With computers now mastering games like checkers, chess, and Go, these games can no longer be the main goals for AI research. Deep Blue's creator, Murray Campbell, said AlphaGo's win was "the end of an era... board games are more or less done."

Compared to Deep Blue or Watson, AlphaGo's methods are more general. This means they could be used for many different problems. Some people believe AlphaGo's success shows that we are getting closer to artificial general intelligence (AGI), where machines can think and learn like humans. This has led to discussions about how society should prepare for such advanced AI.

In China, AlphaGo's success was a big moment. It helped convince the Chinese government to invest much more money in artificial intelligence research.

In 2017, the DeepMind AlphaGo team received a special award called the Marvin Minsky medal. This award recognizes outstanding achievements in AI.

On the Go Community

Go is very popular in China, Japan, and Korea. The 2016 matches were watched by millions of people worldwide. Many top Go players were surprised by AlphaGo's unusual moves. They often looked strange at first but turned out to be brilliant. The Korea Baduk Association gave AlphaGo an honorary 9-dan title for its creative play.

Ke Jie, who was the world's best Go player at the time, first thought he could beat AlphaGo. But after seeing AlphaGo play, he changed his mind, saying it was "highly likely that I (could) lose."

Toby Manning and Hajin Lee, a former professional Go player, believe that in the future, Go players will use computers to help them learn and improve their skills.

After losing the second game, Lee Sedol said he felt "speechless." He apologized for his losses, saying, "I misjudged the capabilities of AlphaGo and felt powerless." He stressed that it was "Lee Se-dol's defeat" and "not a defeat of mankind." Lee said his eventual loss to a machine was "inevitable." However, he also said that "robots will never understand the beauty of the game the same way that we humans do." Lee called his game four victory a "priceless win."

AlphaGo Documentary Film (2016)

The documentary film AlphaGo shows the drama of the games. It focuses on the human Go champion and the AI challenger. The film explores how the human mind works under pressure and how computers make decisions.

Hajin Lee, a former professional Go player, praised the film for showing the feelings and atmosphere of the matches. She noted the close-up shots of Lee Sedol realizing AlphaGo's intelligence and the tension in the room.

James Vincent, a reporter, noted how the film uses emotional cues. He described how Lee Sedol's confidence changed after the first game. Lee became nervous and spent a long time on one move, while AlphaGo responded quickly and calmly. Vincent also highlighted Lee's "divine move" as a moment where humanity struck back.

Murray Shanahan, a professor and DeepMind scientist, said that Go is a great example of what AI can do in other areas. He believes that just as AlphaGo discovered new ways to play Go, AI could help discover new drugs or materials.

Similar Computer Systems

Facebook also worked on its own Go-playing system called darkforest. It also used machine learning and Monte Carlo tree search. While strong against other computer programs, it had not beaten a professional human player by early 2016.

DeepZenGo, another system, lost 2-1 in November 2016 to Go master Cho Chikun.

In 2018, a scientific paper mentioned that AlphaGo's methods could be used to find new pharmaceutical drug molecules. Since then, similar AI systems have been explored for many different uses.

Example Game

AlphaGo Master (white) played against Tang Weixing on December 31, 2016. AlphaGo won by resignation. White's move 36 was highly praised.

First 99 moves

Moves 100–186 (149 at 131, 150 at 130)

Impacts on Go Players

The documentary film AlphaGo made people hope that professional players like Lee Sedol and Fan Hui would learn a lot from playing AlphaGo. However, their rankings didn't change much. In 2019, Lee Sedol announced he was retiring from professional play. He felt he could no longer be the top Go player because of how strong AI had become. Lee called AI "an entity that cannot be defeated."

See Also

In Spanish: AlphaGo para niños