Fat Fritz 2.0 sounds like an update. According to its developer Albert Silver it is by and large a new development. It is a neural network that uses the novel NNUE technology, and it is commercially available from Chessbase. Silver maintains that his creation is “arguably the strongest entity that has ever played chess”. In order to understand its significance we need to delve into the recent history of computer chess.
A few years ago, versions of the open-source programme Stockfish were competing with the commercial engines Komodo and Houdini at the top end of the performance spectrum. In late 2017 DeepMind’s AlphaZero lifted chess to another level. Enough about its technology was made public for it to be copied and developed further. LeelaZero was born, the first open source engine based on a neural network. Leela’s emergence also gave a boost to Stockfish, while their competitors fell behind. Commercial engines seemed to be obsolete. Who would pay for a programme if the two strongest ones are available for free? We recently reported about the free Black Diamond app that brings several engines to the smartphone.
NNUE allows a programme to combine the best of both worlds: The humanly devised position evaluation is replaced by a machine-generated one from a neural network, while the engine does the computing in the traditional way.
The permanent Top Engine Chess Championship has recently been dominated by a duel between two systems: Leela with its revolutionary strategic depth versus the tactically hardly fallible computing monster Stockfish. In an interview with this site, Matthew Sadler explained his fascination with this battle between Leela and Stockfish and the extent to which human chess players can learn from it. The English grandmaster predicted that these two fundamentally different engines would be at each other’s throats for a long time. What he wasn’t aware of in April 2020 was that a fundamental new development had already conquered computer shogi. Since Stockfish also has a powerful shogi engine, it is not a surprise that its developers picked this up.
The NNUE architecture, a reverse acronym for Efficiently Updatable Neural Networks, doesn’t require graphic processors but can run on standard central processing units (CPUs). NNUE allows a programme to combine the best of both worlds: The humanly devised position evaluation is replaced by a machine-generated one from a neural network, while the engine does the computing in the traditional way. Although this makes the programme a little slower, it understands chess much better, and that turned out to be worth much more than the loss of speed. Since no high-performance graphics cards are required, NNUE programmes run on standard computers.
Thanks to an NNUE architecture, Stockfish 12 made a performance leap of about 100 rating points compared to the previous version 11. NNUE Stockfish has since been dominating. In the TCEC 20th season final that ended last week, Stockfish defeated LeelaZero 53–47.
Silver finds the “Zero” attitude purist and unpragmatic. He wondered what results could be achieved if a neural network starts out on a knowledge base of high-class human and computer games.
Fat Fritz 2.0 also runs on an NNUE architecture. Unlike LeelaZero, AlphaZero or its successor MuZero, Fat Fritz hasn’t started from scratch. Albert Silver doesn’t subscribe to the “Zero” attitude of his fellow developers. “Zero” stands for the machine learning on its own through playing against itself, without having any knowledge implanted. The developer from Rio de Janeiro finds this insistence purist and unpragmatic. He wondered what results could be achieved if a neural network starts out on a knowledge base of high-class human and computer games. With this approach he did not make himself exclusively popular in computer chess circles.
“In the course of her learning, Leela went through phases in which she sometimes preferred this opening, sometimes that opening,” Silver told us. This kind of learning has led to Leela limiting herself to learning chess from a limited number of openings and structures. And that in turn means that the engine understands some positions much better than others. For professionals and ambitious amateurs, it is of considerable value to know in which structure they should likely rely on which engine.
Silver started to train his programme in 2019 with world class otb, correspondence and engine games. He also fed it endgame tablebases from the beginning. Originally called Deus X, it became Fat Fritz and was part of Chessbase’s Fritz 17 release in November 2019. When Stockfish NNUE started to dominate thanks to the new NNUE architecture, Silver decided to train a neural network on the basis of the new technology. He was not the only one with the ambition to outperform Stockfish 12. In November 2020, an NNUE Komodo called Dragon appeared. It surpassed LeelaZero in the computer ranking but didn’t quite catch Stockfish 12.
Silver’s neural network is twice as large as that of Stockfish 12. Its positional evaluations are based on the old Fat Fritz. He tested Fat Fritz 2.0 extensively against Stockfish 12. Across the full spectrums of opening it outperformed Stockfish 12 by 45 to 50 rating points, says Silver. He sent Fat Fritz 2.0 to the testers in late November, but as usual the results were only published after the software’s public release this Tuesday. When Stockfish 12 can play the openings prefered by it or chosen by the testers, it is still on top of the Computer Chess Rating List at a superhuman rating of 3566 with Fat Fritz 2 following at 3520. The gap will decrease if the testers run Fat Fritz on four CPUs just like Stockfish, but it will probably not be enough to change the number one spot.
PS: The Stockfish team has released a statement that it is shocked by claims of originality when most of the code of Fat Fritz 2.0 is copied from Stockfish and the modifications are not sufficiently documented as stipulated in the open-source license.
Chessbase published a video interview with Albert Silver.
Fat Fritz 2.0 (requires Windows 7 or higher) €99,90 incl VAT / €83,95 without VAT outside the EU