Well, this is interesting and it really only backs up what I think. I play the 2. ...a6 O'Kelly Sicilian with my own analysed variations, not "book" at all, except that now the Chess.com computer recognises my lines as book whereas two or three years ago it didn't. And if I get 1. e4, I know pretty much where I am and I'm only going to lose via a relatively obvious blunder, whereas d4 lines are more complex and harder to keep track of. Instinctively almost, I think they're stronger that 1. e4, which I see as drawish, and I've played 1. d4 for a couple of decades.
If d4 is going to be played early, I don't believe it can make a negative difference to play it on move one, and if it is not played on move one, logically I think that would be weaker, so I believe your SF will come to favour 1.d4.
An unusual phenomenon
"If d4 is going to be played early, I don't believe it can make a negative difference to play it on move one, and if it is not played on move one, logically I think that would be weaker, so I believe your SF will come to favour 1.d4."
That's the intriguing bit. Does White lose anything from playing what a human would consider a followup move to 1. d4, first, vs playing 1. d4 immediately?
The OP: "What can we learn from this?" Common sense: "Nothing."
"If d4 is going to be played early, I don't believe it can make a negative difference to play it on move one, and if it is not played on move one, logically I think that would be weaker, so I believe your SF will come to favour 1.d4."
That's the intriguing bit. Does White lose anything from playing what a human would consider a followup move to 1. d4, first, vs playing 1. d4 immediately?
For me, the logically correct answer to that is "yes" but only if someone has worked out "best" responses to *all* transpositional moves. But many will argue that after 1. c4, 1. ...e5 is black's strongest move, and of course then normal 1. d4 openings are impossible.
Stockfish 11 recently came out. Claiming to be 50 ELO stronger than SF 10 (currently unproven, SF 10 remains #1 on CCRL ratings), it is the strongest engine available to the public. As a long-time engine admirer and enthusiast, I am always happy to fiddle around with engine matches and analysis in unscientific experiments. The latest of these experiments is underway, with some interesting results.
Chess engines have evolved a lot over the years. When I started playing chess seriously ten years ago, my laptop struggled to reach search depths over 20 moves deep on any engine. These days, a normal computer running Stockfish (an engine relatively better than its peers at reaching deep searches quickly, at the expense of searching broader into shorter lines) can reach 30+ depth without breaking a sweat. My own question here, then, is what does the newest Stockfish believe are the main opening lines from move 0, with no openings book to guide it?
So the experiment began. I turned off all opening theory, set my desktop to infinite analysis on a chess board at move 0, and sat. Currently 55 hours into its analysis.
Some limitations to acknowledge:
The time needed to reach a deeper search grows exponentially, while the yield on skill gain decreases. The largest gain in skill for an engine happens between a search depth of 1 and 2, increasing sequentially as it brute-forces its search until a depth of around 8-12 for most engines, where it then begins alpha-beta pruning (only analyzing lines that are promising for best-play between both sides). It's entirely logical for an engine to become stuck searching through unprofitable lines if it fails to spot the "correct" and/or "better" lines earlier in the analysis. An engine would, in theory, eventually see that its incorrect analysis would lead to a draw or loss and then backtrack to earlier lines, but since draws would often require hitting the 50-move rule (100 moves by search depth count) after a piece is captured, this obviously would take an eternity to analyze.
With all that said, let's finally tackle the question: What does SF 11 think are mainline openings?
And the answer is: it's complicated. Lol.
At first, relatively standard ideas were suggested. Engines always seem to like the French Defense for Black, so 1. e4 e6 showed up a lot. So did 1. d4 Nf6 2. c4 lines.
After reaching 36 moves deep however, strange things started to happen. 1. e4 dropped out of the top two recommendations entirely for a while. The last favored 1. e4 line showed Stockfish switch from the French to the Sicilian:
Stockfish 11 at depth 39.
From there, Stockfish started favoring various 1. d4 lines with an interesting twist: it stopped caring about playing 1. d4 first. Often the first move for White would be variations on 1. e3/Nf3/g3/c4, transposing into more normal positions a few moves later. The current mainlines, at a search depth of 53 moves, ignore 1. e4 lines almost entirely--
What can we learn from this? So far, it appears that the current iteration of Stockfish believes Black's responses to 1. e4 are too powerful for 1. e4 to be the best opening possible. This makes sense in some respect, as Stockfish would have to find refutations to both the Sicilian and the Ruy Lopez, at minimum. However, even though a search depth of 53 (reaching move 26 for Black by standard notation) is deep enough to be out of most human lines, it still isn't truly deep enough to prove anything concrete. Further research is needed, and barring a power outage at my house, it will be interesting to see if the engine's opinion changes in deeper lines.