Chess computer: 25 years of Deep Blue – Digital

Open detailed view

February 10, 1996. Garry Kasparov competes against the chess and super computer “Deep Blue” developed by IBM. Kasparov plays black and loses the first game of the match. He can take the overall victory 4-2, but for the first time a machine has won against the world’s best person in the game of kings. The image of the superiority of the human brain is getting a few scratches. A year later, Kasparov loses the revenge duel against the software. However, the Russian is not convinced of the capabilities of the machine. Deep Blue was no more intelligent than a programmable alarm clock, says Kasparov. The computer only gained through the sheer computing power. He systematically calculated 200 million possible moves per second.

Where does computer chess stand today, 25 years later? Is the game of chess for computer science – as it was called in the 1970s – chess for computer science what the flies of the genus Drosophila are for genetics? Or has the game been solved and has lost its appeal – for computers and also for people?

Johannes Zwanzger knows both sides, he played several years in the chess Bundesliga at SC Forchheim and Bayern Munich. In 2015 the mathematician won the “World Chess Engine Championship”, that is, the world championship in computer chess with his chess program “Jonny”.

“Computer chess programs are now so good that it is an illusion to play against them,” says Zwanzger. Even a world-class player might still do this for training purposes. “The train has left completely for the people.” He also doesn’t play against Jonny himself: “You don’t have a mental arithmetic competition against the calculator either”. When playing a game against a chess program, you usually get the impression that you have no idea about the game and that is frustrating.

Chess computers in the style of Deep Blue do their work using the so-called AlphaBeta algorithm. This is essentially based on brutal arithmetic. This means that the program calculates all game positions up to a certain depth for a position on the board, the so-called game tree. There was also an evaluation function that separates sensible variants from senseless ones. One rule for this is, for example, the evaluation of the characters. “The pawn is worth one point, the knight three, the rook five and the queen nine points. This addition is not always correct, but usually works quite well,” says Zwanzger. Little by little, more and more of these rules were added to the programming of these classic engines.

The basis of the AI ​​are neural networks and an algorithm that calculates the most likely game situations

One of those classic top chess programs is called Stockfish. Stockfish is so far the most successful participant in the World Chess Championship. In 2017, however, the software suffered a devastating defeat – against the artificial intelligence (AI) AlphaZero. AlphaZero won 28 of 100 games, the other games ended in a draw. AlphaZero is software from the Google subsidiary DeepMind. The successor to AlphaGo, who beat the world’s best Go player in 2016, learns through millions of games against himself, all that is given is the rules of the game.

The basis of the AI ​​are neural networks and an algorithm that only calculates the most likely game situations. In addition, the software is rewarded or punished for its “behavior”. This reinforcement learning uses a principle similar to that of the human brain. Kasparov then wrote about AlphaZero that he was happy about the win against Stockfish, because the program, like him, has a dynamic style.

So does AlphaZero play like a human? Computer chess world champion Johannes Zwanzger says: “With Deep Blue I can immediately see from some moves that she was playing a computer.” These are often those that seem pointless at first glance, for example when a figure moves back and forth. For a person it is an admission that you have no idea. AlphaZero, on the other hand, sacrifices a character for an attack. And Johannes Zwanzger’s “Jonny” now also has a neural network. This is integrated differently than with AlphaZero, but has also brought a massive gain in playing strength.

Can you still learn something from computers when you can no longer defeat them? In Zwanzger’s opinion, the ever-improving engines have seen in recent years that you can defend many positions that were previously assumed to lead to a mate. “The pawn moves g2-g4 and especially h2-h4 were still the exception years ago. AlphaZero has now shown that you can be very successful with them, especially in the opening game”.

Another positive effect is that, thanks to the many excellent chess programs, everyone now has the opportunity to get good training opportunities with little effort. However, the computer cannot replace human analysis. “If the computer only throws results without being able to say in a clear form how it came about, then these results are less valuable than the assessment of a strong human player. Objectively, he is certainly less precise, but his reasoning is understandable.”