← StudyX Devlog

How Computers Changed Shogi, Go and Chess

Hiroki ObaraHiroki Obara · DeveloperStudyX MindsportsHow-To
How Computers Changed Shogi, Go and Chess

In May 1997, at the Equitable Center in New York, IBM’s Deep Blue beat the world chess champion Garry Kasparov 3.5 to 2.5 over six games. IBM’s history page “Deep Blue” calls it the first computer system to defeat a reigning world champion under standard tournament time controls.

Within about twenty years the same thing happened in Go and in shogi. This post goes through chess, Go and shogi in that order, with the dates and results taken from the records of the people who built the programs and the people who ran the matches. How the three games differ is in Shogi, Go and Chess: The Differences, and the line of world chess champions up to Kasparov is in World Chess Champions.

Chess: Deep Blue in 1996 and 1997

According to IBM, Deep Blue first played Kasparov in February 1996, a six-game match in Philadelphia. Deep Blue won the first game, the first win by a computer against a reigning world champion under regular time controls. Kasparov then recovered and won the match 4 to 2.

The rematch in May 1997 went like this:

  • Game 1: Kasparov won
  • Game 2: Deep Blue won
  • Games 3 to 5: draws
  • Game 6: Deep Blue won

That made 3.5 to 2.5. IBM writes that Deep Blue ran on 32 processors working in parallel and could evaluate 200 million chess positions per second.

IBM’s page also quotes Kasparov: “I have to pay tribute. The computer is far stronger than anybody expected.”

Go: AlphaGo in 2016

DeepMind’s page “AlphaGo” says that, despite decades of work, the strongest Go programs had only reached the level of human amateurs.

On 27 January 2016, Nature published DeepMind’s paper “Mastering the game of Go with deep neural networks and tree search”. Its abstract opens by saying that Go has long been viewed as the most challenging of classic games for artificial intelligence, because of its enormous search space and the difficulty of evaluating board positions and moves.

The abstract describes two kinds of neural network: “value networks” to evaluate positions and “policy networks” to select moves. They were trained on a mix of human expert games and games the program played against itself. AlphaGo won 99.8% of its games against other Go programs and beat the European Go champion 5 games to 0. The paper calls this the first time a computer program had defeated a human professional player in the full-sized game of Go, something previously thought to be at least a decade away.

DeepMind’s page dates that match to October 2015. The opponent was the reigning three-time European champion.

In March 2016, in Seoul, AlphaGo played Lee Sedol, winner of 18 world titles, and won 4 to 1. DeepMind writes that over 200 million people around the world watched.

Two moves from that match are well known. In game two, AlphaGo played Move 37. In game four, Lee Sedol played Move 78. DeepMind describes each as a move with about a 1 in 10,000 chance of being played. Move 78 became known as “God’s Touch”, and Lee Sedol won that game. The page adds: “Players of all levels have examined these moves ever since.”

Lee Sedol said of Move 37: “I thought AlphaGo was based on probability calculation and that it was merely a machine. But when I saw this move, I changed my mind. Surely, AlphaGo is creative.”

Shogi: the Denou-sen

The Japan Shogi Association (JSA) kept a live blog of the Denou-sen, a series of two-game matches between a professional shogi player and a computer program. In the first Denou-sen in 2016, the program PONANZA beat the Eio title holder in both games.

The second Denou-sen, in 2017, was between Eio Amahiko Sato, who also held the Meijin title, and PONANZA, the winner of the 4th Denou Tournament for computer programs. Each side had five hours. Who would move first was decided at a press event on 22 February by furigoma, a toss of pawns, and the person who threw them was the Go player Lee Sedol. PONANZA got the first move.

  • Game 1 (1 April): PONANZA won in 71 moves
  • Game 2 (20 May): PONANZA won in 94 moves

At the press conference after game 2, Sato said he had lost while putting his own sense of the game and his own values straight up against the program, and that both games were ones in which PONANZA showed a sense of its own, outside his, and was better because of it.

At the same press conference, one of PONANZA’s developers explained how the top programs had grown strong. They now improve through reinforcement learning, studying and improving on their own. They first got that far by learning from tens of thousands of games by professional players, their teachers. He added that he had heard professional players were now studying the knowledge, strategies and feel of computer shogi. The record of this press conference is on the JSA blog page “記者会見(3)” (Press conference, part 3, in Japanese).

Three games, one system: AlphaZero

DeepMind introduced AlphaZero in late 2017. Its post of 6 December 2018, “AlphaZero: Shedding new light on chess, shogi, and Go”, says AlphaZero is a single system that taught itself chess, shogi and Go from scratch. It started from random play and was given no knowledge of the games beyond the basic rules. The full evaluation was published in the journal Science.

According to the post, training took about 9 hours for chess, 12 hours for shogi and 13 days for Go. The results against other programs after training were:

Game Opponent Result
Chess the 2016 world computer chess champion program 155 wins and 6 losses in 1,000 games
Shogi the 2017 CSA world champion version of a shogi program won 91.2% of games
Go DeepMind’s AlphaGo Zero won 61% of games

The same post carries words from Yoshiharu Habu, 9-dan professional. Some of AlphaZero’s moves, such as taking the king to the centre of the board, go against shogi theory and look perilous from a human point of view, yet it stays in control of the board. In his words, its unique playing style shows that there are new possibilities for the game.

And Kasparov again: “Not only do these self-taught expert machines perform incredibly well, but we can actually learn from the new knowledge they produce.”

Timeline

When Game What happened
February 1996 Chess Deep Blue wins one game against Kasparov; Kasparov wins the match 4 to 2
May 1997 Chess Deep Blue beats Kasparov 3.5 to 2.5
October 2015 Go AlphaGo beats the European champion 5 to 0
March 2016 Go AlphaGo beats Lee Sedol 4 to 1
2016 Shogi PONANZA wins both games of the first Denou-sen
April and May 2017 Shogi PONANZA beats Amahiko Sato in both games of the second Denou-sen
Late 2017 Chess, shogi, Go DeepMind introduces AlphaZero

Sources

  • IBM, “Deep Blue”
  • D. Silver et al., “Mastering the game of Go with deep neural networks and tree search”, Nature 529, 484-489 (2016)
  • Google DeepMind, “AlphaGo”
  • Google DeepMind, “AlphaZero: Shedding new light on chess, shogi, and Go” (6 December 2018)
  • Japan Shogi Association, Denou-sen live blog (game reports and press conferences of the first and second Denou-sen, in Japanese)

About Mindsports

Mindsports is an app with 12 mind sports, including shogi, chess, Go, mahjong and bridge, for iPhone, iPad and Android. The beginner lessons for shogi, chess and Go teach the rules through tap-to-answer quizzes and then let you practise in real game positions. The CPU runs from Lv.1 to Lv.50, one level up per win and one down per loss. During a game, Undo takes a move back, and Review lets you step through a finished game one move at a time. A shogi, chess or Go game against the CPU can be left part way and picked up again from the Continue card on the home screen. The games and beginner lessons are free, with ads. A one-time “Unlock All” purchase opens all advanced lessons and removes the ads.

How to use the app is in Getting Started with StudyX Mindsports, and how the CPU level moves is in How the CPU Level Works in Mindsports.

Download on the App Store Get it on Google Play

StudyX MindsportsMind games for families