August 14, 2020

Chess has died, long live chess | Technology

The expression "the king has died, long life to the king" has become popular in some monarchies to dismiss a deceased king, welcome his successor and also avoid the uncertainty that generally lies in the interregnums. The phrase could well be applied also to the history of chess since computers have begun to dispute the reign in the game. Let's see it

The first program to play chess was never executed. He wrote it in the middle of the last century Alan Turing, great mathematician and one of the fathers of computer science and artificial intelligence. It was necessary to wait until the end of the seventies to have computer programs capable of playing chess at a high level. Two decades later, in 1997, Deep Blue, a machine designed by IBM, won the then world chess champion, Gari Kaspárov. The human reign ended there.

The prowess of Deep Blue it was achieved after investing IBM several years in research and many millions of dollars. The strategy followed by this and other computer programs has basically consisted in combining calculation capacity and strategic knowledge in the game. The human knowledge directly injected into the software and extracted from huge repositories of games serves to guide which movements to explore and how to decide which will finally be chosen to move from analysis to action.

The strategy followed by 'Deep Blue' and other computer programs has basically consisted in combining calculation skills and strategic knowledge in the game

At the end of 2018 there was a new milestone in the world of chess, I would say a world landmark, without more appellations. A British company, Deep Mind, acquired by Google in 2014, designed a program called AlphaZero, able to learn to play chess and some other games, like Go, based only on the knowledge of its rules. For this he needs only to play against himself, and to improve his competence as he does so.

After spending a few hours learning how to play, AlphaZero faced Stockfish, the best chess program to date, playing a hundred games. AlphaZero won 28 and tied the rest. Of course, Stockfish played, let's put it that way, with one hand tied behind his back, as his library of openings was limited and the time available for each movement.

In December 2018 the results of a new contest with less restrictive rules of the game were published. AlphaZero again overwhelmed Stockfish. Of a thousand games won 155 and only lost 6, tying the rest.

AlphaZero learns through complex algorithms that are based on the so-called reinforcement learning, common in human learning and other living beings. If a decision is made that is adequate over time, a positive reinforcement is obtained that reaffirms this decision for the future. In the same way, wrong decisions are penalized. This allows the machine to learn without any prior knowledge, beyond, of course, that of the rules of the game and its objective. After learning, AlphaZero operates in general like any chess program, analyzing a large set of possible movements and finally choosing the most promising of them.

We are very far from achieving a machine that has the general learning capacity of a person. Of course, we hope that before we achieve it we have planned how to make it just for our good

But also here there is an important difference with programs like Stockfish. AlphaZero reduces by a thousand times the number of movements explored in each stage by this. Even so, the amount is still much higher than the few hundred movements that an expert in the game usually takes into account before moving one of his pieces, which shows that the human skill in the game is a masterpiece of natural intelligence. In addition, a human expert could suddenly play a variant of chess in which, say, horses instead of moving in "L" moved diagonally, like bishops. However, AlphaZero would have to relearn how to play from the beginning to this new game. We are very far from achieving a machine that has the general learning capacity of a person. Of course, we hope that before we achieve it we have foreseen how to make it only for our good.

One of the results that has most surprised the experts is that AlphaZero has learned strategies that had escaped until now the human development of chess. Perhaps the best way to describe it has been given by the great chess master Peter Heine, by stating that he had always wondered how chess beings of a superior kind, more intelligent than ours, but who already had the answer, would play chess.

Again, a dead king, a king, but in this case with both kings belonging to a new dynasty, that of intelligent machines. We teach the dead king mostly to us to play but the new king of chess is no longer subsidiary of human mastery in the game. He has even learned and taught us things we did not know. Of course, he owes his creation as an amazing machine for learning specific problems and very simple to describe, although enormously complex to solve.

Senén Barro. CiTIUS, Singular IT Center of the University of Santiago de Compostela

Chronicles of the Intangible It is a space for dissemination on computer science, coordinated by the academic society SISTEDES (Society of Software Engineering and Software Development Technologies). The intangible is the non-material part of the computer systems (ie thesoftware), and here they tell their history and their future. The authors are professors from Spanish universities, coordinated by Ricardo Peña Marí (professor at the Complutense University of Madrid) and Macario Polo Usaola (full professor at the University of Castilla-La Mancha).


Source link