16 August 2019

The Value of Deep Learning

Let's take a break from DeepMind's AlphaZero, seen last week in 'Game Changer' PGN, and consider AlphaZero's underlying technology. The motivation is an article that appeared this week, DeepMind's Losses and the Future of Artificial Intelligence (wired.com) by Gary Marcus. It starts,

Alphabet’s DeepMind lost $572 million last year. What does it mean? DeepMind, likely the world’s largest research-focused artificial intelligence operation, is losing a lot of money fast, more than $1 billion in the past three years. DeepMind also has more than $1 billion in debt due in the next 12 months. Does this mean that AI is falling apart?

Author Marcus asks several important questions, of which one touches on chess. Here's the question:-

Is DeepMind on the right track scientifically?

It's a good question, although I suspect it's one of those questions that seven wise men couldn't answer. Here's the chess connection:-

DeepMind has been putting most of its eggs in one basket, a technique known as deep reinforcement learning. That technique combines deep learning, primarily used for recognizing patterns, with reinforcement learning, geared around learning based on reward signals, such as a score in a game or victory or defeat in a game like chess. [...] The trouble is, the technique is very specific to narrow circumstances

While working on posts for this blog, I frequently rely on services like OCR and language translation that have improved considerably over the last five years, mostly thanks to the same technology that was used to develop AlphaZero. How do we put a value on those services?

15 August 2019

2019 CJA Awards - Part 2

I ended 2019 CJA Awards - Part 1 with a promise:-

I'll be back in a few days with the post that I intended for today.

Taking the lead from last year's post, 2018 CJA Awards (August 2018), I'll mention four awards:-

  • Chess Journalist of the Year
  • Best Chess Book
  • Best Chess Art
  • Best Chess Blog

'Chess Journalist of the Year' went to the incomparable Al Lawrence. It's only been three years since he last won the same award, which I covered in 2016 CJA Awards (August 2016):-

The most prestigious of the awards is undoubtedly 'Chess Journalist of the Year', won by Al Lawrence for the second time; the year 2000 was the first (see Chess Life, November 2000).

'Best Chess Book' had two winners:-

  • Instruction: GM Joel Benjamin; Better Thinking, Better Chess
  • Other: GM Andrew Soltis; Tal, Petrosian, Spassky, and Korchnoi: A Chess Multibiography with 207 Games

Last year's 'Best Chess Art' is now split into two categories:-

  • Best Chess Art: Val Bochkov, Melinda Matthews, Natasha Roberts; Chess Adventures with FM Alisa Melekhina, Chess Life Kids, February 2019
  • Best Single Chess Magazine Cover: Joe Jennings, Frankie Butler; Timur Flies High, Chess Life, November 2018

The category 'Best Chess Art' had two honorable mentions, one for Carlotta Notaro and one for Willum Morsch. The artwork behind all three awards can be seen in my previous post 2019 CJA Award Entries (July 2019).

That 'Entries' post also mentioned two candidates for 'Best Chess Blog'. The CJA award went to First Move Chess by John Henderson. In the 'Entries' post I wrote, 'As far as I can tell, the first post in the full blog is dated after the deadline for CJA entries.' This was an error on my part, perhaps provoked by a lack of obvious navigation tools on the blog. The first post appears to have been Advantage Anand (August 2017). Old timers like me can remember Henderson's 'Scotsman' column stretching back to the early days of the web. The chess blogosphere is markedly enhanced by his presence.

With nine award categories and more than 40 subcategories, I've only mentioned a small fraction of the many CJA awards. Congratulations to all winners!

13 August 2019

2019 CJA Awards - Part 1

For this post I intended to follow up last month's post on 2019 CJA Award Entries. Last week the Chess Journalists of America (CJA) published their annual list of awards at Awards 2019 (chessjournalism.org).

The intended post would have looked something like last year's 2018 CJA Awards (August 2018), with a short write-up about the winners in my favorite categories and congratulations to all winners. Unfortunately, the list of 2019 awards is so long and so unstructured that I was compelled to load it into a database to understand it.

Due to a number of technical obstacles that presented themselves along the way, what should normally have been a 5-10 minute task turned into a two hour chore, and I finally ran out of time. While I can't yet say much about specific awards, I can say that there are 64 awards total, categorized as follows:-

43 First Place
20 Honorable Mention
   1 ?

That single '?' is in fact a nearly blank award where only the 'Nominator' is listed. Perhaps there are only 63 awards.

While creating the database, I noticed a few other anomalies. One article, 'The 2019 Denver Open - Chess Life Online May 18, 2019' by IM John Watson, received two first place awards -- one for 'Best Analysis - online', and one for 'Best Tournament Report - State/Local - online'. Another category, 'Best Photojournalism Article', had two first place winners.

I'll be back in a few days with the post that I intended for today.

***

Later: A day after I posted the above, the CJA distributed an email with its list of winners in a more structured format. There's not much to gain by comparing the email to my database, but it did confirm the two separate awards for the same article by IM Watson and a 'tie' in the photojournalism category.

12 August 2019

TCEC S16 L1; CCC9/-10 GPU Blues

In last week's post on the world's top-2 engine tournaments, TCEC S16 L2; CCC9/-10 Endless Bonuses, one event was well underway and the other was marking time. Which was which?

TCEC: 'League 2' is underway and has completed the first of the two round-robins. • CCC: The current tournament, the 'CCC Bonus: CPU Showdown' with five engines, is almost finished.

After a week, the comparison 'well underway' vs. 'marking time' still holds true.

TCEC: 'League 2' finished and 'League 1' started. The following chart shows the top half of the final 'League 2' crosstable.

CCC: A month ago I wrapped up reporting on CCC9 in TCEC S16 Starts; CCC9 Finishes (July 2019). In that post I noted,

It's surprising to see that Leela did not qualify for the [CCC9] final match. I haven't found any Chess.com reports on the series of events, but I'll be patient before looking elsewhere for the reason(s)

Last week Chess.com issued a final report in Stockfish Wins Computer Chess Championship As Neural Networks Play Catch-Up. It summarized CCC9 with:-

Stockfish won CCC9 over Leelenstein, a neural-network chess engine based on Lc0, the leading machine-learning chess project. The champion engine triumphed in the blitz time control of CCC9, beating 17 other engines in a "gauntlet" format.

There was no mention of Leela, other than passing references like 'based on Lc0' in that paragraph. The report gave further details about CCC10:-

Stockfish now looks to defend its title in 'CCC10: Double Digits', an 18-engine tournament played in four rounds. CCC10 is now in progress. The first three rounds of CCC10 will all be played at a time control of 10 minutes plus a three-second increment, on the border between blitz and rapid. The two-engine finals of CCC10 will be 400 games split up among three blitz and rapid time controls to determine the champion.

The declaration that 'CCC10 is now in progress' may have been premature. A few days later, in GPU and CCC10 Status Update, we learned,

I've just been told not to expect the GPU machine to be ready until early next week. That means CCC10 is on hold. Super disappointing, I know, but there's nothing to be done.

That explains the steady stream of bonus events running only CPU engines. Since the GPU engines are the cutting edge of computer chess technology, occasional glitches are to be expected. The filler tournament running now is 'CCC Bonus: Waiting On The GPU'.

[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]

11 August 2019

Mainly for Mathletes?

Some people have way too much time on their hands. I'm talking about the people behind the 129.536 views that this video has notched to date.


30 Weird Chess Algorithms : Elo World (42:35) • 'Published on Jul 15, 2019'

The description explains,

An intricate and lengthy account of several different computer chess topics from my SIGBOVIK 2019 papers. We conduct a tournament of fools with a pile of different weird chess algorithms, ostensibly to quantify how well my other weird program to play color- and piece-blind chess performs. On the way we "learn" about mirrors, arithmetic encoding, perversions of game tree search, spicy oils, and hats.

What's SIGBOVIK? It's The Association for Computational Heresy (sigbovik.org), named after Harry Quokka Bovik. And what about those SIGBOVIK 2019 papers? They're here: tom7 / papers / chess (tom7.org). After that, you're on your own.

09 August 2019

'Game Changer' PGN

Continuing with AlphaZero's Zeros (July 2019), about Sadler & Regan's book 'Game Changer', digging deeper into the book requires the corresponding PGN file. In Chapter 2, titled 'ZeroZeroZero', Sadler wrote,

At the beginning of 2018, I was invited to the DeepMind offices at St Pancras in London to study 210 games from the newest series of matches between Stockfish and AlphaZero. [...] I was provided with two files: a file of 110 games played without an opening book from the starting position, and a file of 100 games starting from various pre-determined opening positions (the positions used in the 2016 TCEC World Championship). The games in each file were grouped by colour so that I first played through AlphaZero’s Black games, and then its White games.

Last December, DeepMind released two files:-

  • alphazero_vs_stockfish.pgn
  • alphazero_vs_stockfish_tcec_positions.pgn

The game counts in each file match the numbers that Sadler gives, meaning that we're probably looking at the same files. Maybe I should say, 'almost the same files', because the two AZ_vs_SF PGN files are not grouped by color as Sadler describes. Where Sadler gives a game number, I also noticed differences.

Another PGN resource is available from newinchess.com (NIC). The product page, Game Changer: AlphaZero's Groundbreaking Chess Strategies and the Promise of AI, says, 'You can download all the games from the book as a PGN-file.' Although the first game in the book -- Kaissa - Chaos, World Computer Championship, Stockholm 1974 -- is missing from the NIC file, the next games in the book match the file. As a bonus, the 'Annotator' tags in the NIC file, e.g.:-

[Annotator "17293012532641473451"]

match the 'Round' tags in the AZ_vs_SF files, e.g.:-

[Round "17293012532641473451"]

This provides a cross-reference between the two sets of files. I'll use these files for further explorations.