AlphaGo Beats The World’s Best Go Player

Published in

Level Up Coding

5 min readNov 16, 2020

Not only is AlphaGo the first machine to defeat a professional human Go player, but the machine is also the first to defeat a Go world-renowned champion, and will always arguably be the strongest Go player in history.

So what is so amazing about a machine beating a human at a board game? Hasn’t this been done before?

Well, to answer those questions, Go is special because it has a reputation for being the most challenging game for artificial intelligence, as a result of its complexity. Prior to DeepMind’s AlphaGo, only the strongest computer programs could play human amateurs at Go. The standard A.I. methods that were previously used to attempt this challenge, were not complex enough to handle the possible Go moves or evaluate the strength of each position on the board.

Go originated in China over 3,000 years ago

What is Go?

Go is not a simple game that anyone can learn by playing it once. Winning this game requires strategic thinking and deep analysis due to its complexity. Two players, one using black and one using white stones, take turns placing the pieces on the board. The end goal is to surround and capture the opposing side’s stones or create spaces of territory. After all the possible moves have been executed, both the stones and empty points are counted, and the person with the most points is the winner. The rules of this game may seem simple, although Go is quite complex. There are an insane 10 to the power of 170 possible board configurations, which is more than the number of atoms that we know of in the universe.

A Brief Overview of how AlphaGo works

AlphaGo has one neural network referred to as the “policy network”, which selects the next move to play. The other neural network is referred to as the “value network”, and it predicts the winner of the game. To train AlphaGo, the machine was introduced to a number of amateur games to help develop an understandable image of what reasonable human play looked like. After, the model played against different versions itself thousands of times to learn from its mistakes. As time passed, using reinforcement learning, AlphaGo improved and became much stronger at learning and decision-making tasks. Even though the search that AlphaGo uses is initially guided based on human-like preferences, AlphaGo can override the bias if the model discovers a move that leads to a better outcome. Soon enough the machine went on to defeat Go world champions in different arenas, earning the title of being the greatest Go player of all time.

A basic diagram demonstrating the policy vs value network.

Winning Against the Strongest Go Player In the World

March 19, 2016, marks the day Google DeepMind’s A.I. program, AlphaGo, beat the strongest Go player in the world, Lee Sedol. The match took place at the Four Seasons Hotel in Seoul’s Gwanghwamun district, marking the start of a big event. The major South Korean television networks broadcasted the game, in China there were 60 million people watching, and on YouTube over 100,000 people tuned in for the stream. Prior to the match against Sedol, AlphaGo played against European Go champion, Fan Hui, beating the champion 5–0. After this match, DeepMind analyzed the results and the algorithm that was able to beat Hui. Contrary to what you might perceive, the performance against Hui did not make AlphaGo seem a likely winner when matched against Sedol, who at the time was the holder of the most Go titles within the past decade. From an estimate, AlphaGo had a 5–10 percent chance of winning each game against Sedol, but this did not stop AlphaGo from taking on the challenge.

“We had the games against Fan Hui as an anchor, where we knew AlphaGo’s strength in October. Going in to this match, we didn’t know its ability now, calibrated against a human player as a strong as Sedol.” -AlphaGo Team

DeepMind was quite surprised that the AlphaGo was able to improve as much as it did when comparing tactics that AlphaGo used against Hui vs Sedol. Even after the successful matches against Fan Hui, people weren’t confident that AlphaGo would win against a Go player as advanced as Sedol, so it was quite revolutionary for the machine to actually beat the world champion at Go.

“I was very surprised, I didn’t expect to lose. I didn’t think AlphaGo would play the game in such a perfect manner.”-Lee Sedol

AlphaGo’s Intuitive Sense Could Change The Future

As a result of Go’s complexity, human players become experts through many years of practice, improving anticipation, decision making, and learning to recognize gameplay patterns. Although the rules to Go are simple and easy to comprehend, the game does have great depth, and the more you play the game, the more you’ll find something new to learn, and the more you’ll feel that you can get better and stronger. As a result of AlphaGo’s learning ability, deep learning allows the model to continually improve its capabilities by playing a large number of games. By doing so the policy network helps predict the next moves, which in return trains the value network to evaluation positions. It is absolutely revolutionary that DeepMind developers have found a way to create an A.I. that has an intuitive sense like AlphaGo, and as this technology advances, it might just be another step forward to developing human-like A.I.

Key Takeaways 🔑

AlphaGo is the only machine in the world that has beat Lee Sedol, the world Go champion
Go is special because it has a reputation for being the most challenging game for artificial intelligence, as a result of its complexity
Prior to DeepMind’s AlphaGo, only the strongest computer programs could play human amateurs at Go
AlphaGo has a policy network, value network, and uses reinforcement learning
As a result of AlphaGo’s learning ability, deep learning allows the model to continually improve its capabilities by playing a large number of games