Here, you can find selected entries from the portal sztucznainteligencja.org.pl. We hope that they will inspire further research on the use of artificial intelligence.

1992: Chinook versus checkers prodigy

It might seem difficult to come up with a better example of a game at which a computer would be better than a human.

The problem is that the number of possible configurations is enormous: 5 x 10 to the power of 20 (500 quintillions).At the end of the 1980s, Jonathan Schaeffer, an IT specialist from the University of Alberta, Canada, decided to build a computer capable of defeating the top checkers champion. He coined the following analogy: let’s imagine that someone emptied the Pacific Ocean and we are supposed to fill it up again... by using a cup. We will need 5 x 10 to the power of 20 of cups.

In 1990, the Chinook computer built by Schaeffer entered the US checkers championship, facing off, for the first time, against Marion Tinsley – the world checkers champion since 1955 (with a hiatus when he relinquished the title).34 games ended in a draw. Tinsley lost two, but won four and so he kept the title.

Two years later, during the world championships in London, Chinook remembered 500 billion possible configurations, but despite that Tinsley managed to draw in six games. Then, stomach ache made it impossible for him to continue and he had to forfeit the game. Diagnosis: pancreatic cancer. The “Beethoven of checkers” died several months later.

Source: https://www.sztucznainteligencja.org.pl/portfolio/1992-chinook-kontra-geniusz-warcabow/

1997: Deep Blue wins against Kasparov

In the seventh move of the decisive game the human made an error in his defense. By sacrificing a knight the machine quickly launched a tremendous attack that forced the human to yield after eleven more moves. “Foul play!” cried the loser.

Cheating? But who would be the cheater? A computer? No. To this day, Garry Kasparov, the chess grandmaster defeated on 11 May 1997 by IBM’s Deep Blue supercomputer, believes that he was facing off not against a machine but another human who was controlling the computer. His proof? According to the Russian the machine’s game was “too human”.

But the world did not believe it was a conspiracy. The fact remained: for the first time in history a machine defeated a world chess grandmaster in a classic game. It happened not only because it could carry out 100 million analyses per second, while Kasparov could only do three. No, the source of Deep Blue’s success was also the application of advanced artificial intelligence techniques that helped it to evaluate its own moves and make only those that seemed the most promising.

Source: https://www.sztucznainteligencja.org.pl/portfolio/1997-deep-blue-zwycieza-kasparowa/

2016: AlphaGo masters Go

“There wasn’t a single moment in which I felt that I was leading,” said Lee Sedol, 18-times world Go champion, when on 15 March 2016, after six days of playing, his loss in a game against a machine was announced. The winner, the AlphaGo algorithm designed by the DeepMind laboratory in London, won the prize of one million dollars and the fame of being the first machine to defeat one of the greatest grandmasters in the history of this ancient game.

How did it happen? First of all, thanks to a deep neural network the machine had memorized tens of millions of moves made by the most experienced players. Then, reinforcement learning had been used: in millions of games, different versions of AlphaGo played against one another, analyzing which moves allowed them to control the biggest territory on the board. In this way AlphaGo had been discovering new strategies.

Finally, the moves from these games had been transferred to a second neural network which trained the system to evaluate potential consequences of each move, to look into the future.

Source: https://www.sztucznainteligencja.org.pl/portfolio/2016-alphago-mistrzem-go/

2017: Libratus, or the art of bluff

You cannot win at poker unless you know how to bluff. That is why for the longest time computers couldn’t compete with humans at this game. However, Libratus changed it all. In January 2017, it defeated four of the best poker players in the world during a 20-day marathon entitled “Brains Vs. Artificial Intelligence: Upping the Ante” at the River Casino in Pittsburgh. While playing, it was learning not from their mistakes, but from the gaps in its own system revealed by its opponents’ tactics.

“The capability of artificial intelligence to reason strategically basing on incomplete data has exceeded the capacity of the best humans,” Tuomas Sandholm, professor of computer science at the Carnegie Mellon University and co-creator of Libratus, commented on the system’s victory. Is this a breakthrough? Yes, it is, and a massive one, too! In all domains where information is incomplete or the opponent uses deception, e.g. in business negotiations, therapy planning, cybersecurity, or military matters.

“One day, your smartphone will negotiate for you the best price for a new car. And that is just the beginning,” said Frank Pfenning, head of the CMU computer science faculty.

Source: https://www.sztucznainteligencja.org.pl/portfolio/2017-libratus-czyli-sztuka-blefu/

2017: AI defeats AI at chess

Google’s AlphaZero defeated Stockfish 8, another piece of software that won the world computer chess championships in 2016.Stockfish 8 lost even though it had access to hundreds of years of human experience in playing chess and decades of experience of other computers, and it could also calculate 70 million positions per second.

In the meantime, AlphaZero could only make 80 000 of such calculations and its creators did not teach it any chess strategies or even standard openings. Instead, it used machine learning to master chess by playing the game against itself.

In the 100 games that AlphaZero played against Stockfish 8 on 17 December 2017 the novice did not lose a single one, but it won 28 and drew 72 times. Reaching such level of play took it four hours.

Source: https://www.sztucznainteligencja.org.pl/portfolio/2017-si-wygrywa-z-si-w-szachy%e2%80%a8/

2019: Rubik’s cube solved in one second

DeepCubeA – a deep neural network created by the scientists from the University of California – solved the Rubik’s cube in 1.2 seconds, nearly three times faster than the fastest of men. On average, it required about 28 moves, while humans usually make 50 moves. At the time of announcing DeepCubeA’s achievement, the current record holder in solving the Rubik’s cube was Yusheng Du. He did it in 3.47 seconds.

DeepCubeA had been trained by using reinforcement learning on 10 billion combinations, thanks to which the AI is capable of finding the fastest cube solving strategies by itself. In the case of the Rubik’s cube there are 43,252,003,274,489,856,000, or more than 43 quintillion possible combinations.

AI algorithms used for solving the Rubik’s cube can be used wherever finding the shortest path to problem solution is required, e.g. in robotics or in anticipating protein structures with which pharmacists would design new medicine.

DeepCubeA’s achievement was announced in „Nature Machine Intelligence” on 15 July 2019.

Source: https://www.sztucznainteligencja.org.pl/portfolio/2019-kostka-rubika-sekunde/