My new book

[Daniel Shawul]

I think Dann has spent hours of CPU time analyzing book positions with different engines. This will definately give you a high quality book, however a particular engine may not like certain openings in the book, unless it is completely neutral that doesn't give an advantage to either side. Dann used to have such a general book, and learning in this case will probably not help (just noise). But most of the time we produce games from large collection of human games that have all sorts of bad variations, in tournaments people cook lines to get a distinct advantage out of the opening, etc.. So reinforcement learning in this case could help the engine to adapt to particular lines in the book that are found to score more.