Do chess AIs still use old-style opening books?

Question

(See also this related question and this related off-site discussion.)

Do state-of-the-art chess AIs still normally start games in human-compiled opening books of master games?

Years ago, as far as I know, state-of-the-art chess AIs started games in opening books. That is, years ago, to choose their first few moves, rather than calculating, chess AIs instead consulted databases of master games.

Do chess AIs still do that?

If you wish to know why I ask, more than 20 years ago, I owned a dedicated chess computer like one of these. The computer would instantly play each of its first several moves, apparently without thinking. The computer would slow down and start thinking only after

• it had exhausted the book or
• its human opponent had made a poor move, deviating from the book.

Nowadays, however, playing on Lichess, where the Stockfish AI is available for post mortem, I notice no obvious sign that Stockfish were consulting anything other than its own analyses.

(One suspects that various interesting things could be written on the topic. My knowledge of chess AIs is limited. I may or may not have posed the question optimally. Therefore, feel free to interpret the question broadly at your discretion.)

Answer 1

Chess engines are, of course, able to analyse opening positions and come up with reasonable moves, but opening books still come with some advantages:

• Moving instantly in the opening saves a lot of time, which can be spent on more complex moves later on in the game.
• Some more subtle theoretical moves in the opening may be missed without an opening book.

In engine vs engine testing, the overwhelming majority of games are conducted starting from an opening position, but the engines themselves are not allowed their own books.

When playing a human, many chess applications use books, but delay their move anyway to give the impression that the program is thinking. This has the advantage of allowing said application to offer a greater variety of openings, whereas relying on the engine's analysis would lead to repetitive games.

In analysis, it is common for to only start from the first move made out of book, because it is likely that moves made in-book are reasonable.

Lichess is somewhat unique in that it uses the engine to analyse every move rather than using a book for the first moves. The analyses of many opening positions are stored in the cloud to avoid duplicating work in the opening (you may see, for example, "Depth 39 [CLOUD]").

Answer 2

You're asking whether AIs still "normally" start games with opening books, and I find it hard to decide what "normally" should be. All engines can be configured to use an opening book.

For analysis purposes like in Lichess, there is obviously no opening book configured as that would mean it wouldn't analyze. Having the engine analyze the position and show an evaluation and a line is what the user wants, not just "5...d6" from an opening book without further information.

In human vs engine play, opening books mostly serve to randomize the openings the engine plays, so there they have a use. But deep opening books that contain huge amounts of lines are not that useful.

In engine vs engine tournament play, the trend is to do without opening books at all because otherwise it's more a battle of the books than a battle of the engines.

So what you get are groups of opening book enthousiasts who play lots of engine vs engine games and constantly try to tweak their books to beat the other engines, that are often the exact same engine but with a book made by someone else. This sort of thing is also useful for correspondence chess -- when a move is known to be a draw from thousands of engines games, it is not that promising for white to play it in correspondence where black can also use an engine.

As for "chess AIs instead consulted databases of master games", that isn't true. They consult an opening book; basically a long list of "in this position, play move X 60% of the time and move Y 40% of the time." Whoever made that opening book probably used a database of games, but the opening book isn't such a database by itself.

Answer 3

There're several closely-related interpretations of your question that can change the answer.

Do chess AIs still use old-style opening books?

As I understand it, all chess engines these days can be configured to use an opening book, or to not use an opening book. So the answer is "if you want them to".

If you want them the engine to use an opening book, then you'll have to supply one. An example is the Cerebellum opening book. You can also, e.g., force the engine to play the Evans Gambit. Some commercial engines will also come with their own opening books.

If you don't want them to use an opening book, then you can disable the book. In this case the engine starts thinking from the first move.

Do state-of-the-art chess AIs still normally start games in human-compiled opening books of master games?

As mentioned, this is a closely-related question that can nonetheless lead to completely different answers because different people use engines differently. For someone who's playing against the engine "for fun", the answer is probably yes, because otherwise the engine only plays a few openings. In engine vs. engine competition, both options happen. The "official" World Computer Chess Championship allows the use of opening books, while the unofficial Top Chess Engine Championship does not. There're advantages to both approaches, the details of which should probably be its own question.

The latter part of this - "human-compiled opening books of master games" - is also debatable. There've been attempts to create engine-compiled opening books, not necessarily of master games, but also including engine vs. engine matches.

Ultimately, engines are programs, and you can configure their settings to do whatever you want.