Is it possible for nowadays strong engines that an engine can write its own opening book? We tried the almighty Stockfish 10. From 82,000 human games between 2600+ elo rated players we extracted from the first 20 moves 1.4 million unique positions and let Stockfish 10 analyze them at 1 second, 2 seconds and 5 seconds. From each we created an opening book.
Then we pitched the 3 books against the handcrafted ProDeo book created by expert Jeroen Noomen over a period of 15 years. None of the 3 Stockfish books was able to win one of the 3x5000 games matches, scores for Stockfish varied from 45-48%, the 45% notable from the 5 second book and so we gave up. Close, but no cigar.
The human still superior in this area?
Encouraged by the excellent results of Lc0 on OKE topping alll lists we tried Lc0 next. The same 1.4 million positions were analyzed at depth=1 (about 100-200ms) which is a factor 25-50 slower than Stockfish and surprise Lc0 was able to win the 5000 game match with a small margin (51.9%).
What does the future hold?
Or deeper analysis depths to begin with.