29 November 2021

Engines Forced to Play Like Us

A few weeks ago, in a post titled TCEC Swiss 2, CCC16 Rapid : Both Underway (November 2021), I wrote,

I certainly wouldn't want to see forced openings used in chess960. It's high time for the chess engine community to investigate a more creative solution to the problem of excessive draws in engine play.

In the traditional start position RNBQKBNR (also known in chess960 as SP518), White has 20 possible first moves. For each of those 20 moves, Black has 20 possible responses. That makes exactly 400 possible positions after one move by each side. Why not run a match in which every game starts with one of those 400 positions, a different position for each game? Since some positions obviously favor one side -- 1.a4 b5 favors White, while 1.a4 e5 favors Black -- the start positions should be played twice, colors switched. That way, no engine has an advantage because of a single dubious position.

Positions well known to current theory, like 1.e4 c5, should be played without any forced book. If 800 games are too many for a match, select the start positions randomly. A 100 game match would need 50 different start positions.

'Oh, no!', purists will say, 'Hundreds of years of experience with 1.d4 and 1.e4 along with millions of recorded games shouldn't be discarded so easily!' That's what people might say. I can't imagine that engines will raise the slightest objection. Since that idea isn't going to fly anytime soon, if ever, let's look at how the openings were chosen for the latest TCEC season.

  • 2021-05-08: Jeroen Noomen and GM Matthew Sadler announce TCEC S21 superfinal book cooperation (chessdom.com) • 'One of the key elements of TCEC is the opening book. Playing at a 3000+ ELO strength, backed by serious hardware, often leads to series of draws in direct battles. Providing imbalanced opening lines is essential to differentiating engine playing strength. Or the way the tournament organizers specify it, “We believe that a champion chess program ought to be able to perform well in a wide variety of opening systems, not merely those that it favors.” The opening book becomes increasingly important as the TCEC championship progresses and is essential in the Superfinal, where the stakes are highest. As of today, the TCEC S21 Superfinal opening book has a serious team upgrade: Jeroen Noomen and GM Matthew Saddler [sic] are teaming up!'

Along with GM Larry Kaufman, GM Sadler is one of the strongest human players participating actively in the world of chess engines. His entry in the Chessprogramming wiki, Matthew Sadler (chessprogramming.org), starts,

English chess grandmaster, chess writer and two-time British Chess Champion, in 1995 and 1997. In 2017, he analyzed the computer games of AlphaZero versus Stockfish, and along with Natasha Regan co-authored the book Game Changer which details the lessons that can be learnt from how AlphaZero plays chess. Since TCEC Season 14 in 2019, Matthew Sadler reports on the Superfinals of the Top Chess Engines Competition, published on TCEC and the ICGA Journal.

I finished coverage of season 14 in Stockfish Wins TCEC S14; CCC6 S2 Underway (February 2019). A companion blog post to the Chessdom article listed the criteria for selecting TCEC S21 openings.

It relates well to a post I recently did on my chess960 blog: The Engine Iceberg Looms Larger (November 2021). I was particularly interested to learn that 'Statistics of the previous two superfinals [TCEC S19 & S20] show that a Leela book exit below +0.30 is an almost 100% certain draw.'

Don't misunderstand me. I'm not questioning the quality of the TCEC S21 preparation. If engines are forced to play like us, this appears to be an excellent approach. I'm questioning why engines are forced to play like us.

Back to the idea for 400 positions, perhaps chosen randomly, the same idea can be extended to chess960. All 960 positions start with 16 possible Pawns moves, 2-4 Knight moves (depending on whether one or both Knights start in the corner), and maximum one castling move (positions patterned '**RK****' or '*****KR*'). That makes between 18 and 21 initial moves per start position, with that quantity squared after one move by both sides.

It's been almost 25 years since IBM's Deep Blue beat Garry Kasparov, then reigning World Champion, in a chess match. Now the engines are rated hundreds of points higher than the best humans. Why do we continue to force the engines to start games the way we do when we are playing at our level? It shouldn't be the only test of chess skill.

No comments: