A few months ago came the news that the company DeepMind, the same that created AlphaGo, a program that defeated the best players of the oriental board game Go, encoded via deep learning and neural networks, had done the same with the theme of chess, creating AlphaZero, that in the words of its creators, in 4 days of self-training and instructing the program only the movement of the pieces, I had achieved a level of International Master. To test such a level, put at AlphaZero to play against Stockfish, to a hundred games, where one of the most strong on the planet could not win a game to AlphaZero.
The experiment was there, and it is clear that AlphaZero does not run on a machine normal, but that probably uses the cloud resources of Google and a powerful infrastructure. However, he was born a project that was called Leela Chess Zero, who apparently plays 2700 Elo points (level chess elite human) in a PC powerful, and that in addition, you can download for free. It is interesting to know that for example, the IBM machine, Deep Blue, had a performance of 12 GFlops and that DeepMind has a computational power of 40 millions of GFlops.
The news of AlphaZero and his triumph over Stockfish were not clear. For example, was tested by making participants do a series of moves by the minute, restricting the possibility of playing the openings with the aid of an electronic book, etc. But perhaps we are ignoring the important thing: it is using a new method, a new technique, to make a program to play a great chess.
After the article of DeepMind, more of a programmer decided to test the idea and thus was born Leela Chess Zero, based on the article mentioned. The system uses the search Monte-carlo and the program resides on the site of LCZero.
LCZero is barely two months active and works in a distributed manner, by asking stakeholders to contribute computer time. The nature of the design of the software of chess use an evaluation function about 80 MB, which grows in strength exponentially as it uses a good graphics card that I have many-core GPUs. Without the card graphics, the performance is only 20% but if you have a machine with four cores (a core i7) with a graphics card like GTX980M, the system plays a 2700 Elo points, which is, as we said, a level of grand master of the elite.
There is a need for measuring the progress of LCZero, which could change the course of chess programs and how we evaluate this game. Probably, and thanks to neural networks, the new way technology seems to be opening up gap.
Illustration courtesy of Chessbase.com