AlphaGo

AlphaGo is a sophisticated computer program designed to play the ancient board game Go. It was developed by DeepMind Technologies, a company based in London that became a subsidiary of Google. Go is considered much more difficult for computers to master than games like chess. This difficulty arises because the game is highly strategic and aesthetic in nature. It also has a massive branching factor, which means there are a huge number of possible moves at any time. Traditional artificial intelligence methods, such as alpha-beta pruning or simple tree traversal, struggle to handle this complexity.

To navigate this complexity, AlphaGo uses a combination of advanced mathematical processes. It relies on an artificial neural network, which is a deep learning method used to process information. This neural network is trained to identify the best possible moves and calculate winning percentages for those moves. AlphaGo also uses a Monte Carlo tree search algorithm to explore different paths of play. The neural network improves the strength of this tree search. This creates a cycle where the program becomes stronger with each new iteration of training. It learns by analyzing extensive data from both human games and computer play.

The history of AlphaGo is marked by several major milestones in artificial intelligence. In October 2015, the original AlphaGo defeated Fan Hui, the European Go champion. This was the first time a computer program beat a professional human player on a full 19x19 board without a handicap. In March 2016, AlphaGo played a historic five-game match against Lee Sedol in Seoul, South Korea. Lee Sedol was a 9-dan professional and one of the best players in the world. AlphaGo won the match with a final score of four games to one. Although Lee Sedol won the fourth game, his victory was a unique moment, as he was the only human to beat AlphaGo in its 74 official games.

During the match with Lee Sedol, AlphaGo utilized significant computing power. Reports from The Economist indicated the system used 1,920 CPUs and 280 GPUs. The match followed Chinese rules and included a 7.5-point komi. In the fourth game, Lee Sedol played a move at position 78 that many professionals called the "divine move." This move caused AlphaGo's computing power to become diverted and confused. A researcher from DeepMind, Aja Huang, later explained that the program's policy network failed to guide it to the correct continuation. This happened because the value network did not recognize Lee's move as highly likely.

Following the Lee Sedol match, DeepMind developed an updated version called AlphaGo Master. This version competed online and in person, eventually playing against the world's number one ranked player, Ke Jie. In May 2017, at the Future of Go Summit in Wuzhen, AlphaGo Master played a three-game match against Ke Jie. AlphaGo Master won all three games, leading the Chinese Weiqi Association to award it a professional 9-dan rank. After these victories, DeepMind retired the AlphaGo program to focus on other areas of artificial intelligence research. The team also released 50 full-length matches as a gift to the Go community.

The evolution of the technology continued with the creation of AlphaGo Zero. This version was unique because it was completely self-taught. It did not learn from any human games or existing data. Instead, it learned entirely by playing against itself. This method proved to be incredibly efficient. AlphaGo Zero surpassed the strength of the version that played Lee Sedol in just three days. It reached the level of AlphaGo Master in 21 days and exceeded all previous versions in 40 days. In a direct match, AlphaGo Zero achieved a 100–0 victory against the earlier competitive version of AlphaGo.

These advancements led to even more generalized artificial intelligence systems. DeepMind researchers developed AlphaZero, which applied the principles of AlphaGo Zero to other games. AlphaZero reached a superhuman level of play in chess, shogi, and Go within just 24 hours. It achieved this by defeating world-champion programs like Stockfish and Elmo. The lineage of this research eventually led to MuZero, a program that can learn how to play games without even being taught the rules. These developments show how machine learning can move from mastering specific games to solving complex, general problems through self-improvement.