Computer Go

Computers can play a game called Go. It is a very hard game. The board is very big. This makes it hard for computers to think.

For a long time, people were better. Even beginners could beat old programs. Some people thought computers could never win.

Then, things changed. New ways of learning helped computers. They learned by playing the game many times.

A program named AlphaGo was very strong. It played against top players. It won many matches in 2016.

Now, computers are the best at Go. They can beat the world's top champions. It is a huge win for science.

Computers can play a game called Go. This game is very hard for machines. The board is big with 361 spots. In chess, players have 20 choices for the first move. In Go, players start with 55 choices. The number of moves grows very fast. This makes it hard for computers to see the future.

For a long time, humans were much better. Old programs were weak. Even beginners could beat them. Some experts thought computers could never win. They thought computers needed to think like humans.

In the late 2000s, a new way helped. It is called Monte Carlo tree search. This way uses many fast, random games to score a position. This helped computers reach an advanced amateur level.

Then, a big change came in 2015. A group called DeepMind used deep learning. This is a way for computers to learn from data. Their program, AlphaGo, was very strong. In 2016, it beat a top pro named Lee Sedol. In 2017, it beat Ke Jie. Now, computers are the best Go players in the world.

Computer Go is a special area of artificial intelligence. It focuses on making computer programs that play the ancient game of Go. This game is much harder for machines than chess. A Go board is very large with 361 intersections. In chess, a player has 20 choices for the first move. In Go, players start with 55 different legal moves. The number of possibilities grows very fast as the game continues. This makes it hard for a computer to look ahead at every possible future.

For many years, computer programs were quite weak at Go. In the 1980s and 1990s, even beginners could beat them. In 1998, strong human players could win even with a 30-stone handicap. A handicap means the human starts with extra stones to make it fair. Even with 15 stones, a program called Go Intellect lost to youth players in 1994. Many researchers thought solving Go was impossible without human-like thinking. Early programs struggled to understand if a group of stones was alive or dead.

Progress began to happen with a new method in the late 2000s. A researcher named Rémi Coulom created the Monte Carlo tree search in 2006. This method works by playing many fast, random games to score a position. These quick games help the computer guess which moves are best. This helped programs like MoGo and Fuego reach an advanced amateur level. By 2011, a program called Zen reached a high rank on the KGS server. These tools allowed computers to finally play like skilled amateurs.

Everything changed in 2015 thanks to a group called DeepMind. They used deep learning to create a program named AlphaGo. Deep learning is a way for computers to learn from large amounts of data. In 2016, AlphaGo played a famous match against Lee Sedol. Lee Sedol was a 9-dan professional player. AlphaGo won the match and became the first program to beat a top professional without a handicap. In 2017, AlphaGo also defeated the world number one player, Ke Jie.

Today, the field of Computer Go is very different than it was before. Many teams now build programs using the same deep learning ideas. Some programs, like Leela Zero, are open for anyone to use. We can see how these machines have conquered the greatest human champions. This journey shows how much computers can learn when they use new ways to think. It is a huge leap from the weak programs of the 1980s.

Computer Go is a specialized field of artificial intelligence (AI) research. Its goal is to create computer programs capable of playing the traditional board game Go. While games like chess were mastered by machines in the late 1990s, Go has remained a much harder challenge. Many researchers believed that defeating top human players was impossible without creating human-like intelligence. This difficulty arises because Go requires intense strategic thinking and intuition. The game is much more complex than chess due to the size of the board and the number of possible moves.

The primary difficulty stems from the massive scale of the 19x19 Go board. This board contains 361 intersections where stones can be placed. In chess, a player has only 20 possible opening moves. In Go, a player begins with 55 distinct legal moves. As the game progresses, the number of branching possibilities grows incredibly fast. This makes it nearly impossible for traditional algorithms to look far enough ahead. Algorithms like alpha-beta minimax, which worked well for chess, often fail on Go. They cannot process the vast number of potential future moves within a reasonable time.

Another major hurdle is the task of board evaluation. To play well, a program must assess which player is currently winning. In chess, computers can count material, such as how many pieces a player has left. In Go, evaluating a position is much more abstract. A program must determine if a group of stones is "alive" or "dead." This involves complex analysis of how stones connect and influence the board. A stone might not seem important immediately, but it can become vital much later. If a program evaluates a position incorrectly, it will make poor strategic choices.

For decades, computer Go programs were quite weak. In the 1980s and 1990s, even beginners could defeat them. In 1998, very strong players could win while giving the computer a 25-to-30 stone handicap. This is an enormous advantage that few humans would ever accept. In 1994, a program called Go Intellect lost all three games in a championship against youth players, even with a 15-stone handicap. Early researchers tried to use human-like expert knowledge to help these programs. However, these "knowledge-based" systems often had major weaknesses in their overall strategy. They could handle specific local situations but lacked a global understanding of the game.

Progress changed in 2006 when Rémi Coulom introduced the Monte Carlo tree search. This algorithm works by creating a tree of potential future moves. Instead of calculating every move perfectly, the computer performs many fast, random playouts. These playouts act as a way to "score" the different paths in the tree. While random play might seem inaccurate, the tree search helps correct the results. This method allowed programs like MoGo and Fuego to improve significantly. By 2009, programs could finally reach the "low-dan" level, which is the rank of an advanced amateur.

The most significant breakthrough came with the deep learning era. In 2015, the research group DeepMind developed AlphaGo. This program combined Monte Carlo tree search with deep learning, a method where computers learn from massive amounts of data. In 2016, AlphaGo played a historic match against Lee Sedol, a 9-dan professional. This was the first time a top professional played a computer without a handicap. AlphaGo won the match, marking a massive milestone in AI history.

DeepMind continued to push the boundaries of what was possible. In 2017, AlphaGo defeated Ke Jie, who was then the world's number one ranked player. Later that year, DeepMind revealed a version of AlphaGo that learned entirely through self-play. This version was even stronger, winning 89 out of 100 games against the previous version. Since then, many other teams have built high-level programs by following these techniques. Projects like Leela Zero and Tencent's Fine Art have also reached professional levels. The field has moved from weak, beginner-level tools to machines that can conquer the world's greatest champions.