Russell Sherwood

Russell Sherwood  Saturday, April 17, 2021

I have been doing a lot of chess coding recently but more on that another time. One tool I have developed is to allow the review of ongoing events and to test it I focused on the Counties and Districts Championships of this season.

Running my prediction routine gives the following expected Finishing League Table….

3 of the Top 5 teams are Welsh and Lancashire both benefitted from a very favourable draw and have a strong Welsh Contingent!

On this basis we can expect the two Essex Teams to be fighting over avoiding the 2nd relegation spot!

More ominous is that the East Wales B team appears to be on a promotion trajectory from the 2nd Division as well!

Russell Sherwood  Thursday, March 11, 2021

Our first skunkworks project was the development of Eman's Experience book.   For those not aware, this is a database where lots of information is stored about positions: Analysis score, Games Percentages, ongoing evaluation changes and much more. This database is then used by the engine (which is an aggressive Stockfish derivative) to enhance its own search.

Over the last few months a small group of us have quietly worked on this and now have a database of 15million moves, covering 13.5 million positions.  Other people have larger databases, but the difference is that ours are all at a depth which aids the engine is correspondence analysis.

If you want to gain access to this you need to be a premium or international member and then would be  able to join the group, although anyone wanting to  needs to contribute to the project, which is fairly simple.

More on the next skunkworks project soon!

Russell Sherwood  Wednesday, March 10, 2021

https://www.iccf.com/event?id=87609

This is the first of several articles reviewing our progress in Friendly and International Events. 

First Up with have the Wales v England 960 Training match. This event was organised to give a number of English players who did not make their National squad a game and the Welsh 960 team (with others) some 960 practice.

It was always likely to be a one-sided match due the difference in playing strength, with a target score-line of around 70% for Wales. The results are turning out to be even better than that with the current score being 22-1 in Wales favour and an expected final score of around 26-2!

I would use the PGN viewer to share a game, but it tends to not get on with 960 games!

Russell Sherwood  Tuesday, February 16, 2021

Disclaimer: Any opinions given are my own and not necessarily representative of any group or body I am associated with.

I had intended to give a run down on the latest in Engine development but instead, based on the number of contacts I have had over the weekend, will focus on one – Fat Fritz 2

Fat Fritz 2 was recently released in a swathe of advertising by Chessbase. They claimed this engine was the new number 1 and so on……https://en.chessbase.com/post/fat-fritz-2-best-of-both-worlds

Without going into detail, this release caused great consternation within the Stockfish and wider engine development community, over copyleft (yes left – look it up!), naming, charges and claims.

I will not go into much depth on the Copyleft issue but in simple terms, under the Stockfish GPL License the full code, including the network if embedded, has to be published. Initially the engine was published without the network, then with a different network to that being sold. 

Whilst this may or may not be still breaking the rules, it is still causing some major arguments.  In the defence of Chessbase  there are several other Stockfish variants which refuse to publish source code – Eman, Zeus, AI and Killifish come to mind, although the key difference is that none of these engines are Commerical

The reality is that the original part of Fat Fritz 2 is a larger network and (less originally) that this network is based on Neural Net Evaluations rather than Stockfish. From this, the issue caused is that the engine should be really called Stockfish + Fat Fritz 2 Network.

Putting all this aside (We are here to play Correspondence Chess after all) – how strong is Fat Fritz 2 and is it worth buying?

Chessbase claim this is the new number 1 – having beaten Stockfish 12 is a match. There are a number of issues with this claim……

1. The match size was tiny – to prove actual decisive superiority requires a much larger match with hundreds or preferably thousands of games involved.
2. Stockfish 12 was used , whereas FF2 is based on a Stockfish Dev(1). This would mean that the engine, ignoring the network, would be around 30 elo stronger. In practice, we are comparing apples with oranges!

This means that the results of the match are inconclusive and not really indicative of a clear new number 1.  Independent testing has shown that FF2 is around the same strength as Stockfish rather than being any stronger. For example https://www.sp-cc.de/

To complicate matters there are two versions of Fat Fritz – The Commercial Version and the Stockfish Github Version. The network released with the Free version is around 45 elo weaker than the commercial one.

Practically the Correspondence player is not really interested in the results of those matches as they are played at, for us, very rapid time controls. I undertook some soft testing of the Free version over the weekend – running it on a few threads alongside my main engine. Most of the time it suggested the same moves but on occasion it did provide some interesting alternatives. 

It is worth commenting that the similar results are  seen if Stockfish and LC0 are run in parallel.

Looking from a wider perspective: Is this engine likely to be stronger for Correspondence Chess. I suspect the answer is yes, as the more knowledge embedded in the network, the better the overall moves should be. The same is true for LC0 networks – a well-trained large network should give better suggestions than a smaller well-trained network, if time is not a factor.

Is it worth the money – my gut feel is if, you want an edge for 1-2 month and consider 80 euros a reasonable price then yes, otherwise use the Free version (WCCF members get in touch if you want a copy) to generate ideas and alternative lines.

Over the next few months, I would expect that:

1. Unless Chessbase release regular updates, FF2 will become weaker than Stockfish.
2. New networks, based on either the CB Network, Free Network or developed from scratch will appear., which will be stronger than the FF2 Commercial network.
3. Other network structures for NNUE will be developed which may eclipse the current standard or FF2 structures.

Of course Ceres may also upset this apple cart but more on that next time.

(1)As a reminder Stockfish do a major release every so often – For example Stockfish 11 and Stockfish 12. In between these major releases minor improvements are released, almost daily. These minor releases (Stockfish Dev) tend to add an elo or two to Stockfish at a time. When enough of an overall gain is made, a new Major release is made

Russell Sherwood  Thursday, January 7, 2021

Here we will get an obvious one out the way. I initially thought of this as hardware but, on reflection, it is not as simple as that.

Chess Engines dominate modern cc, and it is assumed that the player with the bigger hardware will prevail and upgrading hardware is a path to success. However, it is not as simple as this.

Any computer has a given level of processing power, so upgrading hardware will increase processing power. However, what really matters is overall computing resource. In its most simple sense this is processing power x analysis time. Sin theory improved analysis can be achieved by increased processing power or increase analysis time. 

In practice this equation would have a number of other inputs, including effectiveness of analysis and creativity of analysis, but we will look at these at a later point.

So, we can increase our processing power as a path to winning but what tends to happen is that is can become an “arms race” with other players with diminishing returns. 

For players, unable or unwilling, to increase their computing resource the option of increasing their Analysis time is an option but this comes with its own set of issues, primarily the significant reduction in the number of games which can be played concurrently. 

So what hardware should the player look to obtain. This is a surprisingly difficult question to answer and much depends on your choice of supporting engine and analysis methods. For a period of time High-end graphics cards were a must as LC0 achieved a level of dominance. Then when Stockfish NNUE came on the scene, CPUs became more relevant again. Now Ceres is coming the needs to change again.

For the aspiring player there are a few key choices

Intel or Ryzen (as a general rule Intel chips will have fewer but more powerful threads) – which of these is most useful depends on your approach to analysis – which will be covered in a future article.

As far as a high-end Graphics card goes – much depends on your view of the value of MCTS engines. Personally I believe they are worth having, if you can afford them. In not then alternatives such as Fat Fritz and online LC0 offering exist.

Russell Sherwood  Thursday, January 7, 2021

This is the first of a series of articles around winning in correspondence chess. It will almost certainly contradict itself and meander at times!

As you climb the ratings ladder in correspondence chess, winning becomes much harder, relatively speaking that “over the board” chess. Some of these reasons are related to the increasing strength of chess engines but many other reasons are also relevant.

 It is the intention of these articles to provide “food for thought” rather than a blueprint!

Russell Sherwood  Friday, January 1, 2021

I've been asked by a few players about what events they can enter - below is a list of standard events, with entry based solely on rating. Other qualifications for these events based on a combination of rating & Title can and do exist but are not shown here.

Russell Sherwood  Sunday, November 22, 2020

A quick tip here. If you use these engines consider the following:

• Make sure you have the latest GeForce drivers and GeForce Experience utility. Within  this is a performance tuning tool. When used this has generated reporting performance increases from 10-20% Nodes per second.
•  

• The second are more important tip is to ensure that you set the RAMLimitmb,. If not set, then LCZero can however up all your available RAM and slow the system down (and it will be using much more than the amount it shows you on the taskmaster!). What should it be set to? Experimentation will be the key here. They key below shown in default install in Aquarium is set to 1000mb (1 Gb) as a starting point. From here I would increase upwards to the point the system started to slow down.