Fat Fritz 1.1 update and a small gift – Chessbase News

Posted: March 8, 2020 at 10:47 am

without comments

3/5/2020 As promised in the announcement of the release of Fat Fritz, the first update to the neural network has been released, stronger and more mature, and with it comes the brand new smaller and faster Fat Fritz for CPU neural network which will produce quality play even on a pure CPU setup. If you leave it analyzing the start position, it will say it likes the Sicilian Najdorf, which says a lot about its natural style. Read on to find out more!

If you havent yet updated your copy of Fat Fritz, now is the time to do it as it brings more thanminor enhancements or a few bug fixes. This update will bring the first major update to the Fat Fritz neural network, stronger than ever, as well as a new smaller one that is quite strong on a GPU, but also shines on even a plain CPU setup.

When you open Fritz 17, presuming you have Fat Fritz installed, you will be greeted with a message in the bottom right corner of your screen advising you there is an update available for Fat Fritz.

When you see this click on 'Update Fat Fritz'

Then you will be greeted with the update pane, and just need to click Next to get to it

When Fat Fritz was released with Fritz 17, updates were promised with the assurance it was still improving. Internally the version number of the release was v226, while this newest one is v471.

While thorough testing is always a challenge since resources are limited, a match against Leela 42850 at 1600 nodes per move over 1000 games yielded a positive result:

Score of Fat Fritz 471k vs Leela 42850: +260 -153 =587 [0.553] Elo difference: 37.32 +/- 13.79

1000 of 1000 games finished.

Also, in a match of 254 games at 3m +1s against Stockfish 11 in AlphaZero ratio conditions, this new version also came ahead by roughly 10 Elo.

Still, it isnt about Elo and never was, and the result is merely to say that you should enjoy strong competitive analysis. For one thing, it is eminently clear that while both Leela and Fat Fritz enjoy much of the same AlphaZero heritage,there are also distinct differences in style.

Perhaps one of the most obvious ways to highlight this is just the start position. If you let the engine run for a couple of minutes on decent hardware, it will tell you what it thinks is the best line of play for both White and Black based on its understanding of chess.

As such, I ran Leela 42850 with its core settings to see what it thought. After 2 million nodes it was adamant that perfect chess should take both players down the highly respected Berlin Defence of the Ruy Lopez.

Leela 42850 analysis:

info depth 19 seldepth 56 time 32675 nodes 2181544 score cp 23 hashfull 210 nps 75740 tbhits 0 pv e2e4 e7e5 g1f3 b8c6 f1b5 g8f6 e1g1 f6e4 d2d4 e4d6 b5c6 d7c6 d4e5 d6f5 d1d8 e8d8 h2h3

This is fine, but it is also very much a matter of taste.

Fat Fritz has a different outlook on chess as has already been pointed out in the past. At first it too will show a preference for the Ruy Lopez, though not the Berlin, but given a bit more time by 2.6 million nodes it will declare the best opening per its understanding of chess and calculations is the Sicillian Najdorf.

Within a couple of minutes this is its mainline:

info depth 16 seldepth 59 time 143945 nodes 7673855 score cp 28 wdl 380 336 284 hashfull 508 nps 54227 tbhits 0 pv e2e4 c7c5 g1f3 d7d6 b1c3 g8f6 d2d4 c5d4 f3d4 a7a6 f1e2 e7e5 d4b3 f8e7 e1g1 c8e6 c1e3 e8g8 f1e1 b8c6 h2h3 h7h6 e2f3 a8c8 d1d2 c6b8 a2a4 f6h7 a1d1 b8d7 f3e2 h7f6

From a purely analytical point of view it is quite interesting that it found 10.Re1! in the mainline. In a position where white scores 52.5% on average it picks a move that scores 58.3% / 58.9%.

Remember there is no right or wrong here, but it does help show the natural inclinations of each of these neural networks.

Even if chess is ultimately a draw, that doesnt mean there is only onepath, so while all roads may lead to Rome, they dont all need to pass through New Jersey.

Trying to find the ideal recipe of parameters for an engine can be daunting, and previously multiple attempts had been made with the well-know tuner called CLOP by Remi Coulom. Very recently a completely new tuner 'Bayes-Skopt' was designed byKarlson Pfannschmidt, a PhD student in Machine Learning in Paderborn University inGermany, who goes by the online nickname "Kiudee" (pronounced like the letters Q-D). It was used to find new improved values for Leela, which are now the new defaults.

His tuner is described as "A fully Bayesian implementation of sequential model-based optimization", a mouthful I know, and was set up with his kind help as it ran for over a week. It produces quite fascinating graphical imagery with its updated values. Here is what the final version looked like:

These values, slightly rounded, have been added as the new de facto defaults for Fat Fritz.

This is a completely new neural network trained from Fat Fritz games, but in a much smaller frame. Objectively it is not as strong as Fat Fritz, but it will run much faster, and above all it has the virtue of being quite decent on even a pure CPU machine. It wont challenge the likes of Stockfish, so lets get that out of the way, but in testing on quad-core machines (i.e. my i7 laptop) it defeats Fritz 16 by a healthy margin.

Note that this is not in the product description, soneedless to say, it is more nor less a gift to Fritz 17 owners.

Enjoy it!

More stories on Fat Fritz and Fritz 17...

See the original post:

Fat Fritz 1.1 update and a small gift - Chessbase News

Written by admin |

March 8th, 2020 at 10:47 am

Posted in Alphazero