Fat Fritz 2 es más fuerte que Stockfish

por Conrad Schormann
16/02/2021 – La contienda, de cuál es el mejor módulo de ajedrez el martes 9 de febrero ha comenzado de nuevo. "Fat Fritz 2.0" ha salido a la venta. Aunque suena a una versión en plan de actualización, Albert Silver, el autor del programa, insiste en que se trata de una novedad absoluta. Según las series de pruebas que ha realizado Silver, el Fat Fritz 2.0 debe ser unos 45 - 50 puntos Elo más fuerte que módulo más fuerte de antes, el Stockfish 12. Antes de que AlphaZero revolusionara el ajedrez (informático), durante mucho tiempo las versiones de Stockfish competían por el máximo rendimiento con los programas comerciales Komodo y Houdini, teniendo los tres más o menos la misma fuerza de juego. Entonces Deep Mind desarrolló AlphaZero, que jugaba al ajedrez como si fuese de otro planeta. Pero se trataba de un proyecto experimental que no estaba disponible para el gran público. Eso sí, la tecnología servía para copiarla y perfeccionarla. El artículo completo por Conrad Schormann, en inglés.

Fat Fritz 2 Fat Fritz 2

Fat Fritz 2.0 incluye la versión más reciente de la interfaz de usuario más avanzada en el mundo con funciones de entrenamiento, análisis automáticos y una base de datos con 1 millón de partidas. Asimismo, está incluida la membresía para la Cuenta ChessBase Premium durante seis meses.


Más fuerte que Stockfish

El artículo original por Conrad Schormann (en alemán), publicado en Perlen vom Bodensee, 08.02.2021

El artículo en inglés

The battle over which is the best chess engine reignites on Tuesday, when "Fat Fritz 2.0" goes on sale. Contrary to the name, which may sound like an update, it is a new development, according to developer Albert Silver. And according to Silver's test series, it is 45 to 50 Elo stronger than the previously dominant Stockfish 12.

Before AlphaZero turned (computer) chess on its head, Stockfish versions were competing with the commercial programs Komodo and Houdini at the top of the performance spectrum. Then the DeepMind development AlphaZero presented chess from another star system. The project was not accessible to the public, but the technology could be copied and further developed.

LeelaZero emerged, the first chess engine based on a neural network. Leela's emergence also gave a boost to the Stockfish development, while the two commercial competitors fell behind and became more or less obsolete (why buy a chess program when the two best are available for free?). From then on, the TCEC computer chess world championship was dominated by the battle of the systems: Leela with its revolutionary strategic depth against the tactically hardly fallible computing monster Stockfish.

Battle of the systems

In an interview with this site, grandmaster Matthew Sadler once explained his fascination with this battle between Leela and Stockfish - and the extent to which human chess players can learn from it.

Sadler predicted that these two fundamentally different engines would be at each other's throats for a long time. What he didn't know was that another fundamental new development had already conquered computer shogi and was long on the way to transforming computer chess as well: the NNUE technique is based on combining the best of both worlds: the human devised position evaluation is replaced by a machine-generated neural network, while the engine does the computing in the traditional way.

This makes the engine a little slower than before, but on the other hand it understands chess much better, and that is obviously worth much, more than the loss of speed. For the userthere is also a practical advantage: unlike pure neural networks, NNUE engines run on traditional CPUs and no high-performance graphics card is required.

Thanks to NNUE technology, Stockfish 12 made a performance leap of about 100 Elo compared to the previous version 11, a milestone. Since then, NNUE Stockfish has dominated computer chess. In the most recent TCEC final in early 2021, Stockfish defeated LeelaZero 53-47.

The computer chess expert Albert Silver from Brazil, who is closely connected to ChessBase, was sceptical about the "zero" attitude of the AlphaZero and LeelaZero developers from the beginning. "Zero" stands for the machine learning on its own in the game against itself, without having knowledge implanted. Silver finds the insistence on the Zero approach dogmatic and purist. He asked whether it might not lead to even better results to give the neural networks a fund of knowledge of high-class games already played right from the start. With this, he did not make himself extensively popular in computer chess circles.

"In the course of her learning, Leela went through phases in which it sometimes preferred this opening, sometimes that opening," Silver explains in an interview with this site. This kind of learning has led to Leela limiting itself 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 which engine they should most likely rely on in which structure.

"Tens of thousands of hours of computer power"

Silver trained his own neural network at the end of 2019. It wasn't supposed to start from scratch, but was fed top-class human and engine games as well as tablebases, right from the the beginning. Fat Fritz was born – and actually went to the top of the computer chess engine rankings in early 2020. This list was only reshuffled at the end of the third quarter of 2020, when the new NNUE Stockfish 12 appeared and was henceforth the boss in the ring of 64 squares. The new technology surpassed everything that had gone before.

At this stage Silver started to train a neural network based on the new technique that would make it possible to beat Stockfish 12. And he was not the only one with such ambitions. Back in November 2020, an NNUE Komodo called Dragon appeared, which is currently ranked second in the engine rankings (behind Stockfish 12, ahead of LeelaZero).

Tens of thousands of hours of computer power and thousands of dollars have gone into training and developing a new neural network, Silver tells us. By the end of November 2020 the program was ready. The operators of computer chess lists have had it since then, and it has long been tested in comparison with other engines. "The results may only be published when the software is released," says Silver.

The technology behind Fat Fritz 2.0 is based on that of Stockfish 12, but it is a completely different programme. The neural network of its engine is twice as large as that of Stockfish 12, the engine "knows" twice as much. Moreover, the basis for the positional evaluations is different, it is completely based on the first Fat Fritz.


Conrad Schormann, periodista alemán