olzpoll.blogg.se

However, it is in some ways tiresome imprinting on one’s own neural network knowledge about opening variations. Fat Fritz and LCZero are already beginning to change opening theory.Įvery average human brain is light years ahead of neural networks when it comes to mastering everyday situations. Nevertheless, here for the rst time in many years we can record a real breakthrough in chess programming. With a painfully practical limitation: Fat Fritz needs (like LCZero) a very high performance Nvidia graphics card (“GPU”) in order to achieve its full playing strength. The moves suggested in analysis are often extremely human and planned. As things stand, Fat Fritz defeats in a direct comparison all traditional chess programs and even LCZero. The result is so convincing that we are now publishing it as “Fat Fritz” along with Fritz17. This approach was followed logically by our longserving technical editor Albert Silver and based on the LCZero technology he trained a neural network for a whole year with GM games. The idea soon came to use our existing base of hundreds of thousands of good grandmaster games to shorten this learning process. LCZero too follows the Google philosophy, that the neural network only learns from games played against itself. Suddenly a chess engine was available whose different analysis results provided new ideas on all fronts. The Open-Source- Project LCZero began to retrace the trail blazed by Google and in the meantime has acquired considerable strength. Nobody had expected that a cooperative effort by chess developers would soon make this technology generally available. Fascinating because it was possible to hope that one could learn really new stuff about chess from this radical approach.

Sobering in the sense that the decades old tradition of chess programming had been relegated to the shadows by a self-learning system. This news was sobering and fascinating at the same time. In December 2017, a press release from Google shook the chess world to the core: its subsidiary Deep Mind built a neural network, dubbed Alpha Zero, which "learned" chess solely by playing millions of games against itself, yet was strong enough to beat Stockfish 8, a leading chess engine.