December's book in the book-a-month project is Bayesian Optimisation by Peng Liu. You won't have seen it in the photo I took for the original post where I said I'd start working through my books in a semi-meaningful way, because I got the book after that post. I succumbed to a mailshot from Springer offering a selection of books for £15 each - falling into some kind of trap of thinking that they must be good value at £15 each when the list price is much higher (£55 in this case). Perhaps Springer also knows that I see books as positional goods, too, and knew I would crumble.
Well, I felt like learning more about Bayesian Optimisation would be a good thing, and hence bought the book and started reading. Setting aside the intellectual quality of the book for a moment, what struck me is that it is a very badly designed and made book. All the £15 books I got were sent to a print-on-demand service, and the quality is a bit variable. One arrived with the pages not properly cut. The Liu book, did not suffer that, but it very much gives the impression that it was not designed to exist in hard copy. The margins are ragged, there are mis-sized figures with poor resolution, and a lot of blank space at the bottom of pages since a figure follows on the next page and there wasn't enough room to put it at the bottom of the partly-empty pages. All things that would not happen in a professionally-designed book. Still, ultimately it is the intellectual content that is important. How did that shape up?
Well, I have not read the whole thing, but read through the first half, starting with the first chapter giving an overview of the topic, chapter 2 on Gaussian processes and chapter 3 on "Bayesian decision theory and expected improvement". I think it is fair to say that I got the gist of chapter 1, made sense of little bits of chapter 2, and found chapter 3 incomprehensible. I'm not even sure how to describe why I didn't understand it, because I don't even have enough understanding to do that. In the end it just felt like word spaghetti as every noun as it moved through the book seemed to pick up more adjectives "eventual multistep lookahead marginal gain". I think if there were (more) specific examples to work through, it would have helped. Chapter 2 did have one, although it was a rather abstract example, and that did aid understanding a bit.
I can't help but feel I am at fault - too dim - to understand the book, and that may be the case, but it's also possible that I should be able to make sense of a book about this kind of technical subject given my prior experience. If only I had learned anything from it, I might be able to calculate which of these two positions is the most likely.
Anyway. I'm afraid I can't recommend it, though I may revisit if and when I learn the topic from some other source to see if I can retrospectively make sense of it from the position of already knowing the material.
Some pictures:
 |
| Me holding the book |
 |
| An example of empty space because a figure is coming up on the next page |
No comments:
Post a Comment