30 November 2020

A Mind-boggling Tactical Battle

No one can accuse me of overusing an easy idea. It's been ten months since I last featured a video on this weekly engine series -- see Stockfish vs. Leela Explained (February 2020).


Stockfish 12 Wins TCEC Season 19! (17:19) • '[Published on] Nov 3, 2020'

The description of the current video said,

Join NM Sam Copeland as he breaks down this epic matchup between computer chess engines Stockfish 12 NNUE and Leela Chess Zero during the TCEC Season 19 Superfinals!

Sam Copeland? Where have we seen that name before? The first reference was 2018 CJA Awards (August 2018):-

For my favorite category, 'Best Chess Blog', the winner was not a blog, but a single blog post: 'Playing The Quintessential American Tournament: The 2017 World Open' (chess.com) by Sam Copeland.

The most recent reference was earlier this month in Komodo NNUE (November 2020):-

We learned, 'Komodo Releases Powerful New "Dragon" Chess Engine' (chess.com; by Sam Copeland).

The game featured in the video is one of those mind-boggling tactical battles where the winner sacrifices heavy material for an enduring attack. I don't remember if the video mentioned it, but the game was no.18 in the 100 game match that I documented in Stockfish Wins TCEC S19; CCC15 Still Halted (October 2020). To play through the moves, see Stockfish 202009282242_nn-baeb9ef2d183 vs. LCZero v0.26.3-rc1_T60.SV.JH.92-190 - TCEC - Live Computer Chess Broadcast (tcec-chess.com/archive). That is one long title! Good thing I wrote the post about Parsing Engine Names (October 2020).

It's also a good thing that Chess.com is allowing resources to be used for covering the TCEC, a rival to its own Computer Chess Championship (CCC). For more Sam Copeland videos, see youtube.com/c/chess/search?query=copeland.

26 November 2020

Queen's Gambit Staying Power

Last month's Yahoo post Queen's Gambit Mania (October 2020) -- where 'Yahoo' is a code word for chess in the mainstream press -- started,

Of the 100 chess stories that appear at the end of every month on Google News, three or four on the same topic will signal strong interest in that topic from the mainstream press. What to say of 15 stories on the same topic?

Another month has passed and the mainstream interest in the blockbuster Netflix series continues unabated. It's possibly the biggest mainstream chess story since Bobby Fischer died in 2008. While working on the 'Mania' post, I captured the Google 'Full Coverage' page shown below at the end of this post.

It goes without saying that the chess press coverage was even more maniacal than the mainstream news coverage. Kudos for the most intense chess press interest go to Chessbase. I counted ten stories, many of them with user comments, to which I could have added a few more stories:-

As for mainstream news sources, The Guardian tops the list of stories that I noted during the past month.

The Netflix series appeared unexpectedly on my radar at one other time during the past month. Near the beginning of the month, one of my pages zoomed to first on the 'Top 10 Pages' for my domain, eclipsing the perennial leader, Index to the World Chess Championship (m-w.com). The new leader, a page I wrote in 2006, titled Top 10 Myths About Chess (ditto; 'People say the darnedest things about chess'), is frequently in the top-10, but rarely leads it. At first I was baffled, but then I noticed the last myth was:-

10. Women can't play chess as well as men • To date it is true that women have not performed as well as men in chess events. There are many possible reasons for this. One may be that male players are often expert at making female players feel uncomfortable at chess events. The Polgar sisters have gone a long way to convince the chess world that women can play very well. Perhaps one day we will discover that women can even play better than men. No one really knows.

Thanks Beth Harmon, Netflix, and Walter Tevis, for making us all think in new ways about our ancient game.

***

Google News 'Full Coverage', end-October 2020

23 November 2020

TCEC FRC2 Underway; CCC 'Currently Uncertain'

Two weeks ago, in Stockfish Wins TCEC Cup 7; CCC GPUs Back, I reported,

TCEC: In TCEC Cup 7, Stockfish confirmed its S19 domination by beating LCZero in the final match. The site is currently running 'Sufi Bonus 3'. • CCC: The GPU problems are apparently over. The site is currently running an event called 'Checkmate in 4' with a GPU engine, Lc0. Before that it ran an event called 'Mystery Engine Debut and Tuning'. Will CCC15 be resumed anytime soon or will it be abandoned?

Last week's post was about the 'Mystery Engine', Komodo NNUE, aka 'Dragon'. What has happened on the two main engine vs. engine sites in the last fortnight?

TCEC: After 'Sufi Bonus 3', the site ran a chess960 event, dubbed 'FRC2'. It started with 16 engines in four 'Leagues' (A to D), followed by eight engines in two 'Semileagues' (1 to 2), followed by four engines in a 'Final League', followed by two engines in a 'Final' match. The 'Final League' is currently underway.

CCC: The site is currently running the 'CCC Blitz Championship 2020' with 16 engines, scheduled to finish in a few days. Stockfish appears to have an insurmountable lead over a group of five engines bunched within a two point range, three of them GPU engines. What's next? According to the site's !Command, '!next = Currently uncertain'. The page Work At Chess.com (chess.com) currently has a listing for:-

Computer Chess Championship Tournament Director • Do you love chess engines? Are you a fan of the Computer Chess Championship? Do you know how to run cutechess and configure engines on servers? If so, send us a message telling us about you and your background to: [email address]

Former TD Greco hasn't been seen on Chess.com since early October.

[For further information from the various stakeholders in the engine-to-engine events, see the tab 'TCEC/CCC Links' at the top of this page. • NB: Leela = LC0 = LCzero]

22 November 2020

Hopelessly Entangled Topics

Another post on the Queen's Gambit Mania (October 2020) would be a great choice for this month's continuation of the series on The Sociology of Chess (November 2016). Since I already used the subject in this month's video selection, Queen's Gambit on the Tomatometer, let's go with a different idea. There's only one problem: it's not exactly a new idea either. The subject has already been seen on this blog several times this year, including a previous sociology post of the month, Political Pundits and Pogchamps (August 2020).


Online chess makes its biggest move (5:09) • '[Published on] Nov 1, 2020'

The video clip is from the CBS Sunday Morning show. Its description says,

The pandemic has hurt most professional sports, but it's done wonders for one game: chess! Players and fans, looking for competitive thrills, are logging into online chess platforms like never before. Correspondent David Pogue talks with a chess grandmaster and an online chess star about the boom in the game.

The chess grandmaster is Maurice Ashley, the focus of last month's sociology post, Wild Rabbit, Meet Black Bear (October 2020). The online chess star is Alexandra Botez, the focus of a previous month's sociology post, The Dilemma of Women's Chess (July 2020).

Which of the two themes -- Netflix 'Queen's Gambit' -or- the online boom -- will turn out to be the biggest chess story of 2020? I vote for Netflix. The online boom still has to prove it won't turn into a bust, like so many previous chess booms.

***

I had a good runner-up video for this post, World's hardest "Chess puzzle" | Fun Puzzles | Chess piece placement puzzle, from Youtube's MathsSmart channel.

Over a hundred years ago, Kempelen's famous chess player could not only beat most players that challenged him, but he could also formulate chess puzzles that stumped the best minds of the day. Here is one of the hardest. You are required to place four Black Queens and a Black Bishop on a chessboard so that they control the entire board. In other words, after the five pieces have been positioned, it will be impossible to place the White King on any vacant square without being in check.

The solution is available from the video page. Math and chess will always be hopelessly entangled.

19 November 2020

2020 Candidates - Early World Championship Steps

I ended a post last week, Karjakin/Carlsen's Early World Championship Steps, saying,

While we're waiting for the '2020 Candidates Tournament, Yekaterinburg' to resume, I might take the time to determine how all eight of the current candidates first qualified into a World Championship cycle.

To do that, I took the names of the eight players from my page on the 2020 Candidates Tournament, then looked up each name in the Index of Players to find the first World Championship event for that player. The following table summarizes the results of that exercise:-

Alekseenko K 2019 WCup
Caruana F 2009 WCup
Ding Liren 2011 WCup
Giri A 2013 WCup
Grischuk A 2000 KO
Nepomniachtchi I 2011 WCup
Vachier Lagrave M 2009 WCup
Wang Hao 2005 WCup


The 'KO = Knockout' format was the same as the 'WCup = World Cup'. For each of the KO/WCup events, I looked up the qualifying paths ('QP') in the 'Clippings' section at the bottom of the Index of Zonals. The following overview summarizes the result for each of the six unique events listed in the table above.

2000 KOZonal Qualifiers 2000-2001 (C19)

Qualified as Junior
[...]
8. A. GRISCHUK RUS 2601

2005 WCupZonal Qualifiers 2004-2005 (C22)

110,8602883,"Wang, Hao",,"CHN",,,"Zone 3.5, 2005",

2009 WCupZonal Qualifiers 2008-2009 (C24)

e) 46 players from European Championships 2008 & 2009
34. M. Vachier-Lagrave (FRA) 2008
[]
h) 6 nominees of the FIDE President
122. Caruana, Fabiano (ITA)

2011 WCupZonal Qualifiers 2010-2011 (C25)

e) 46 players from European Championships 2010 & 2011
29. I. Nepomniachtchi (RUS) 2010
[]
i) 6 nominees of the FIDE President
122. Ding Liren (CHN)

2013 WCupZonal Qualifiers 2012-2013 (C26)

d) From FIDE Rating List, average 3/2012 up to 1/2013
24. A. Giri (NED) 2715,33

2019 WCupZonal Qualifiers 2018-2019 (C29)

Alekseenko K [?]

Only for the last event (2019 WCup), did the relevant QP page fail to identify the qualifying path for the player. According to Wikipedia's page on Kirill Alekseenko:-

Although [Alekseenko] failed to qualify for the Chess World Cup 2019 through European Championship placement, he was chosen as a wildcard nominee by the organiser of the tournament.

Indeed, my page 'Zonal Qualifiers 2018-2019' has a number of blanks for the various nominees:-

Nominees of the FIDE President
1. Kovalenko Igor (LAT)
2. Salem A. R. Saleh (UAE)
3. Urkedal Frode Olav Olsen (NOR)
4. Ziska Helgi Dam (FAI)
5.
Nominees of the Organizer
1. Iljiushenok Ilia (RUS)
2. Pridorozhni Aleksei (RUS)
3.
4.

The exercise was useful not only to learn more about the importance of the different qualifying paths, but also to identify some weaknesses in my own pages. • To do: (1) Fill in those three blanks for the nominees to '2019 WCup'. (2) Crosslink (in both directions) the IZ/KO/WCup pages with the corresponding QP page for faster navigation between them.

16 November 2020

Komodo NNUE

In last week's post on chess engines, Stockfish Wins TCEC Cup 7; CCC GPUs Back, I mentioned a CCC event called 'Mystery Engine Debut and Tuning'. This led to the question:-

What was the 'Mystery Engine'? We may never know, although some sources are certain it was 'Komodo NNUE'.

When I wrote that, I reckoned that 'We may never know' was likely to be short lived. Indeed, a few days after the post we learned, Komodo Releases Powerful New 'Dragon' Chess Engine (chess.com; by Sam Copeland). It started,

The Komodo team has released Komodo 14.1 and a new chess engine dubbed "Dragon." Dragon adds powerful NNUE (Efficiently Updatable Neural Networks) technology to the successful Komodo engine.

After a short explanation of the NNUE technology...

NNUE engines have proven to be a major advancement in chess engine technology, allowing traditional tactically powerful "brute force" engines to incorporate the deeper positional understanding evinced by neural network engines such as AlphaZero and Leela Chess Zero.

...the article confirmed that Komodo NNUE was the mystery engine:-

In a debut and tuning tournament in the Chess.com Computer Chess Championship, Dragon finished fourth behind Stockfish 12, Leelenstein and Leela Chess Zero.

That sentence was followed by the crosstable for the 'Mystery Engine' tournament. Along with some marketing hype for the Komodo engine -- multiple 'World Champion' (i.e. ICGA) titles -- the article included a link to a previous article by the same author about Stockfish NNUE, Stockfish Absorbs NNUE, Claims 100 Elo Point Improvement (chess.com; August 2020), which I don't think I had seen before.

15 November 2020

Chess Movie 1927/1938

Since the year 2020 is the year of the chess movie -- a few months we had Chess Movie 2020 (August 2020; 'Critical Thinking' by John Leguizamo), and more recently Queen's Gambit Mania (October 2020) -- it's easy to forget the old classics. Let's go back more than 80 years to a chess movie of yesteryear.


"Le joueur d’échecs" (1938) © Flickr user manhhai under Creative Commons.

The Flickr description, given first in Vietnamese then in English, is a copy/paste of the introduction to Wikipedia's The Chess Player (1938 film):-

The Chess Player is a 1938 French historical drama film directed by Jean Dréville, released by Compagnie Française Cinématographique in France and Columbia Pictures in the United States and starring Françoise Rosay, Conrad Veidt and Bernard Lancret. It is a remake of the 1927 silent film The Chess Player.

A link in the description leads to Le joueur d’échecs (1938; conradveidt.wordpress.com; 'The Conrad Veidt Website'), with 25 photos from the film, including the poster shown above. For the IMDb writeups on the two versions of the film, see:-

  • The Chess Player (1927; imdb.com); • 'In 1776, an inventor conceals a Polish nobleman in his chess-playing automaton, a machine whose fame leads it to the court of the Russian empress.'

  • Chess Player (1938; ditto) • 'A toymaker in Poland specializes in building lifesize mechanical men. He builds a chess-playing "automaton" to hide a pretty young Polish activist who is being hunted by occupying Russian forces.'

The red text in the image above is not on the original poster. It is a watermark that says, 'Scanned by Monique classique'.

12 November 2020

Karjakin/Carlsen's Early World Championship Steps

Yesterday, in a post titled Opens for Discussion on my World Chess Championship Blog, I wrote,

The Association of Chess Professionals (ACP) recently issued a Proposal For The World Championship Cycle, where the essence of the proposal was: "It seems natural then to make the open tournaments part of the World Championship cycle."

While I was researching the topic, I wondered which early steps the current World Champion had taken on his path to the World Championship. A few years ago, in Karjakin/Carlsen's World Championship Careers (May 2016), I started a similar analysis for the two players who were preparing for a title match later that year.

In a curious coincidence of parallel paths, I discovered that both players had launched their careers at the 2004 FIDE Knockout Matches, Tripoli, and were both eliminated in the first round. The next step for both was the 2005 World Cup, Khanty-Mansiysk, the first in a new series of World Cup events which carried on the tradition of the knockout format.

Since writing that post, I've added a few more pages to my history of the World Championship. The page 2002-2004 Zonal Cycle Qualifiers says that both players qualified the same way into the 2004 Knockout event:-

106. Karjakin, Sergey (UKR) – President’s nominee
109. Carlsen, Magnus (NOR) – President’s nominee

Continuing with the parallel path, a similar page, 2004-2005 Zonal Cycle Qualifiers, says that both qualified via the same event into the 2005 World Cup:-

42, 1503014, Carlsen Magnus, g, NOR, Warsaw 2005
49, 14109603, Karjakin Sergey, g, UKR, Warsaw 2005

The official name for Warsaw 2005 was the '6th European Individual Chess Championship'. It is one of the events mentioned on my page of clippings for the 2004-2005 Zonal Cycle. Karjakin finished a full point ahead of Carlsen, although Carlsen was more successful in the 2005 World Cup. According to the post on 'Karjakin/Carlsen's Careers':-

Karjakin was again eliminated in the first round. Carlsen survived to the fourth round, where he lost the match but continued to play for a qualifying place in the subsequent candidates matches. He succeeded, but in the first match was eliminated from the rest of the cycle by Aronian.

That rupture of the parallelism was the first sign that Carlsen would eventually be more successful than Karjakin on the path to become the World Champion. While we're waiting for the 2020 Candidates Tournament, Yekaterinburg to resume, I might take the time to determine how all eight of the current candidates first qualified into a World Championship cycle.

09 November 2020

Stockfish Wins TCEC Cup 7; CCC GPUs Back

In the previous post on the world's two foremost engine-vs-engine competitions, Stockfish Wins TCEC S19; CCC15 Still Halted (October 2020), I reported,

TCEC: Stockfish beat LCZero in the S19 final by a final score of +18-9=73. The match was technically decided after 93 games. • CCC: The title 'CCC15 Still Halted' said it all, although there was a short, unconfirmed rumor that the 'GPUs are back'.

In the meantime, the TCEC has conducted its premier side event, while the CCC shows signs of resuming normal operation.

TCEC: In TCEC Cup 7, Stockfish confirmed its S19 domination by beating LCZero in the final match. Stoofvlees beat Komodo in the consolation match for 3rd and 4th places. The last time I reported on a TCEC Cup, Alliestein Wins TCEC Cup 6; CCC Tests 'Stockfish NNUE' (August 2020), the main takeaway was 'In Cup 6, the AI/NN engines dominated.' How could we have known that NNUE would breathe new life into Stockfish ?

The site is currently running 'Sufi Bonus 3'. The '!next' plan says, 'next FRC2 testing and FRC2 ~1.5 weeks'. When was FRC1? As far as I can tell, it was more than six years ago. In Stockfish, the Strong (July 2014; Chessdom.com: 'Stockfish is TCEC double champion'), I started the post asking 'Which engine is the strongest at chess960?' Then I went on to feature the event in two posts on my chess960 blog:-

I'm looking forward to reporting on FRC2 for the same blog.

CCC: The GPU problems are apparently over. The site is currently running an event called 'Checkmate in 4' with a GPU engine, Lc0. Why is it called 'Checkmate in 4'? We may never know. The associated 'Info' tab lists the four engines that are participating, says 'Testing and tuning new engine', and gives an associated PGN file. I downloaded the file and discovered that it was for an event called 'Mystery Engine Debut and Tuning', which was the name of the event preceding 'Checkmate in 4'.

What was the 'Mystery Engine'? We may never know, although some sources are certain it was 'Komodo NNUE'. Whatever the mystery engine was/is, it finished fourth out of seven engines, including Lc0 plus two other GPU engines.

Will CCC15 be resumed anytime soon or will it be abandoned? For the first report on the CCC15 GPU problems, see TCEC S19 DivP Chugging Along; CCC15 Halted (September 2020).

[For further information from the various stakeholders in the engine-to-engine events, see the tab 'TCEC/CCC Links' at the top of this page. • NB: Leela = LC0 = LCzero]

08 November 2020

Queen's Gambit on the Tomatometer

Given the current wave of Queen's Gambit Mania (October 2020; 'The trailer has almost 1.4 million views and 2500 comments'), I wasn't surprised to see that half of my picks on the short list for this month's video were also about the Netflix series titled 'The Queen's Gambit'.


How Anya Taylor-Joy Became Beth Harmon in 'The Queen's Gambit' | Rotten Tomatoes (9:04) • '[Published on] Oct 20, 2020'

After reading the video's description by Rotten Tomatoes...

How do you turn yourself into the world’s most formidable chess prodigy when you barely know how to play the game yourself? That was the question facing Anya Taylor-Joy as she prepared to step into the mind of Beth Harmon, the protagonist of Walter Tevis's novel 'The Queen’s Gambit' and its lavish Netflix adaptation, which arrives on the service this month.

Set in the 1950s and '60s, the series follows Harmon’s rise from orphan with a talent to teen prodigy to national champion, all as she deals with addiction to tranquilizers -- crippling, but sometimes useful when she’s plotting her moves alone at night -- and the men of the chess-world boys' club who want to keep her out.

...you might think you don't have to watch the series or read the book, but there's much more to the story than that. The description continues by explaining that the clip is about the back story:-

Ahead of the series’ release, Taylor-Joy sat down with us to break down how she tapped into Harmon’s pain, why she relates to her solitary pursuit, and how her dance training helped her feign a mastery of pawns, rooks, and knights.

For more from Rotten Tomatoes, see The Queen's Gambit (rottentomatoes.com). There's nothing rotten here. The series gets 100% on the Tomatometer from the critics and 97% on the Audience Score.

03 November 2020

November 1970 & 1995 'On the Cover'

It's been one year since the post November 1969 & 1994 'On the Cover' (November 2019) announced,

In November 1969, Chess Life and Chess Review were merged for the first issue of Chess Life & Review. Let's continue the 'On the Cover' series with CL&R from 50 years ago, together with CL from 25 years ago.

The 1970 CL&R had 60 pages with less than six pages of 'Tournament Life' announcements. The 1995 CL had 84 pages with 17 pages of tournament announcements. The 6th American Open (1970, Santa Monica), the premier tournament held over the long Thanksgiving weekend, had a $25 'Regular' on-site entry fee and a $1000 first prize. The 31st American Open (1995, Los Angeles) had a $129 on-site entry fee in the top section, with a $3300 first prize.


Left: 'Andy Soltis (foreground) and Ken Rogoff, Members of the World Champion Student Team'
Right: 'Dato' Tan: World's Richest Amateur?!; Botvinnik Remembered...'

Chess Life & Review (50 Years Ago)

The headline and byline for the featured event said, 'Haifa 1970 : U.S. Student Team Captures World Title by Frank Skoff and Dick Verber'. The first paragraph said,

With a last-minute spurt, the U.S. Student Team captured 1st place in the 17th Annual World Student Team Championship, held this summer at Israel's beautiful Mediterranean Seaport, Haifa. England led from the very start, and it was not until the final round that we nosed them out in a dramatic finish. Eleven countries participated; Russia, and her Eastern allies, chose to boycott the tournament, protesting Israel's "imperialist aggression" against her Arab neighbors.

The Skoff/Verber report placed Rogoff on first board for the U.S. team, Soltis on second board. Today both Andrew Soltis and Kenneth Rogoff are prominent ambassadors of the chess community -- Soltis as a world class historian and writer, Rogoff as a GM celebrity thanks to his status as one of the world's leading economists.

Chess Life (25 Years Ago)

The upper right corner of the CL cover said, 'Inside: Official U.S. Chess Holiday Catalog'. The 'On the Cover' writeup opposite the masthead had two parts. Both related to the part of the world bordering the Pacific Ocean:-

Dato' Tan may arguably be the richest man in the world of chess organization and sponsorship. While most of his largesse is evidenced on the Pacific Rim, he played a large part in securing sponsorship for the U.S. Amateur Team playoff at the Harbor View Marina and Yacht Club in Baltimore, in 1993.

Grandmaster Utut Adianto won the Pacific Zonal, which was held in Malaysia. Scoring 9-2, a full point and a half ahead of the rest of the field, is certainly noteworthy, but the event captured the attention of the world press for another reason. Ian Murray of the Queensland Chess Association issued a statement on behalf of the players, protesting France's resumption of nuclear testing in the Pacific. The statement called for all chessplayers to boycott the use of the French Defense ...

Boycott the French Defense? Tongue in cheek?

As for ex-World Champion Mikhail Botvinnik, who had passed away 5 May, the issue had a three page writeup by ex-Soviet refugee and American GM Leonid Shamkovich (translated by GM Alex Fishbein), followed by a one page appreciation by Anthony Saidy.

02 November 2020

Stockfish NNUE - Three Useful Pins

A month ago, in Stockfish NNUE - Three Early Threads (October 2020), I tried to understand the structure of the technology behind Stockfish NNUE, but ran into some basic questions. It turns out that some of my questions are answered in the Stockfish Discord pinned messages, i.e. posts that members of Stockfish Discord have flagged as particularly useful.

Many of the Discord pins are intended for Stockfish developers who need guidance on how to access various elements of the underlying software. Other pins explain the theory behind NNUE. All three of the following messages were posted in July 2020 by nodchip, who is identified by the Chessprogramming wiki page Stockfish NNUE (chessprogramming.org) as the father of the concept:-

Stockfish NNUE [is] a Stockfish branch by Hisayori Noda, aka Nodchip.

First, I wondered whether the board representation included both Kings. Indeed it does:-

A standard input feature is called HalfKP. "K" is the one-hot encoding [NB: a group of bits with a single '1'; the rest are '0'] of the position of a King. "P" is the one-hot encoding of the position and the type of a piece other than the Kings. Note that "P" consists of the friend's pieces and the opponent's pieces. "P" also contains a feature representing that a piece is captured, and removed from a game. This is "+1". The number of the dimensions of "P" is 10 * 64 + 1 = 641. HalfKP is the direct product of "K" and "P". The input feature vector represents a board state.

NNUE holds two HalfKP vectors both of White's King and the other pieces, and Black's King and the other pieces. It inputs the White's input feature vector into the upper half of the input layer, and the Black's into the lower half of the input layer in a White's turn. It inputs the Black's into the upper half of the input layer, and the White's into the lower half of the input layer.

I also wondered about the apparent inefficiency of the encoding. It is related to the 'UE' portion of the NNUE acronym, 'Updatable Efficiently'.

If we calculate the output of the neural network with a simple way, we need to execute the calculation above for each layer. It takes a long time, and the nps [NB: nodes per second] will be extremely low.

At this time, we can speed up the calculation by using a special characteristics of the feature vector. When a move is played in a position, only a few elements in the corresponding feature vector are changed. For example, if a Pawn is advanced, the element in the feature vector corresponding to the previous position of the Pawn will be set to 0, and the element corresponding to the current position will be set to 1. If we calculate only about the changed elements, we can speed up the calculation. This technique is called "difference calculation" in computer shogi.

Difference calculation can be applied only for the calculation between the input layer and the first hidden layer. Because the input vector of the second and later hidden layer are changed drastically when a move is played. But this is fine because the number of the network parameters between the input layer and the first hidden layer is very large, and the others are very small.

We can not use difference calculation if a king is moved in HalfKP. Because all the elements in the feature vector corresponding to the king will be changed.

Finally, much of the early Discord discussion centered on selecting the positions for training. This assumes some familiarity with neural network training, but will make sense to anyone who knows the basics.

Training data: The training data are FENs [NB: chess positions] generated from games where Stockfish plays itself at fixed depth or nodes. The recommended move, the searched score, and outcome of the game are recorded in each position.

Training: The network is trained on these positions, and aims to output the searched score and the predicted output of the game. (Note that the predicted outcome can be converted to/from the searched score.) As it trains, it automatically updates the network parameters after each iteration. When the training is finished because the loss cannot be lowered, you get your network.

Evaluation: In a general neural network, all the output of each layer are computed everytime. But NNUE only calculates the difference between the previous position and the current positon to calculate the output of the first hidden layer. This drastically speeds up the calculation because NNUE has almost all the network parameters between the input layer and the first hidden layer, which is why it is called an efficiently updatable neural network. That is what makes it so interesting.

Together that makes three important pieces of the NNUE implementation: the structure of the input positions, the reason for that structure, and how known positions are used. I think I'm finally starting to grasp the NNUE concept. It's only taken three and a half months!

01 November 2020

Chess as Caprice

In last month's post, French or Italian?, the most recent post in the long-running, monthly series on Top eBay Chess Items by Price (March 2010), I wrote,

According to my informal selection rules, I should pick an item from the previous month, i.e. September 2020, but the item shown below ended on 1 October and none of the alternatives on my short list were particularly interesting. I'm also hoping that the rest of October will offer other items for the next eBay post.

Hoping sometimes works. The item pictured below was titled 'William Gropper b.1897-d.1977 etching Chess Players Blacklisted shop Studio' and sold for 'US $1,195.00 or Best Offer'. So which was it, $1,195.00 or Best Offer? I don't remember the final eBay price being so ambiguous, but maybe I haven't been paying sufficient attention.

The description said only,

William Victor "Bill" Gropper, was a U.S. cartoonist, painter, lithographer, and muralist. A committed radical, Gropper is best known for the political work which he contributed to such left wing publications as The Revolutionary Age, The Liberator, The New Masses, The Worker, and The Morning Freiheit.

That description happens to be the first paragraph of the artist's Wikipedia page, William Gropper. But what about the word in the title, 'blacklisted', or the phrase in the description, 'committed radical'. What's that all about? Wikipedia continues,

His parents were Jewish immigrants from Romania and Ukraine, who were both employed in the city's garment industry, living in poverty on New York's Lower East Side. His mother worked hard sewing piecework at home. Harry Gropper, Bill's father, was university-educated and fluent in 8 languages, but was unable to find employment in America in a field for which he was suited. This failure of the American economic system to make proper use of his father's talents doubtlessly contributed to William Gropper's lifelong antipathy toward capitalism.

Gropper's alienation was accentuated when on March 24, 1911, he lost a favorite aunt in the Triangle Shirtwaist Factory fire, a disaster which resulted from locked doors and non-existent exits in a New York sweatshop. Some 146 workers burned or jumped to their deaths on that day in what was New York's greatest human catastrophe prior to the terrorist attacks of Sept. 11, 2001. [...]

Due to his involvement with radical politics in the 1920s and 1930s, Gropper was called before the House Un-American Activities Committee in 1953. The experience provided inspirational fodder for a series of fifty lithographs entitled the Caprichos.

'The Caprichos', aka 'The Capriccios', in English 'The Caprices', are also described in Wikipedia, Los caprichos, although in a different context:-

Los caprichos are a set of 80 prints in aquatint and etching created by the Spanish artist Francisco Goya in 1797 and 1798, and published as an album in 1799. The prints were an artistic experiment: a medium for Goya's condemnation of the universal follies and foolishness in the Spanish society in which he lived.

A century and a half later, Goya's Caprichos/Capriccios were an inspiration for Gropper. Is this month's chess etching one of 'The Caprichos'? My guess is 'No, it isn't', but it's certainly a caprice, at least for some people.