shahbazc@gmail.com (4) [Avatar] Offline
#1
I'm looking forward to this book and I'm assuming that the goal here is to delve into reinforcement-learning, unlike almost all other ML/DL books. I personally care about the game of Go only because it will help me learn RL.

I noticed that the 88 pages published so far (which I quickly scanned) are just building up to neural networks. My personal preference is that very basic knowledge of deep learning should be assumed (since manning has several books which introduce the subject). I don't mind a review, but I hope the book is mostly about reinforcement learning, rather than deep learning with RL being relegated to one or two chapters.

I've had trouble understanding RL so far because the few other books out there over-explain relatively simple to understand concepts such as MDPs but jump right over the mish-mash of equations for q-learning, utility calculaiton, etc. These equations often have recursion as well as iteration and although they probably look simple to someone who already understands them, they are quite difficult to understand initially. RL becomes even more confusing because authors very quickly go over various iterations: value iteration, policy iteration, sarsa, table blah blah that it leaves a newbies head spinning!

I realize this book will show how to use deep learning + RL but I wouldn't mind getting some historical context for how RL was used before deep learning.

Looking forward to learning more about RL, I personally find it to be MUCH more interesting than other branches of ML, and just as confusing!
KevinF (17) [Avatar] Offline
#2
Yes, RL is a major focus of the book, and we are dedicating nearly a third of the book to it. I agree with you, I find it one of the coolest fields in ML. We are a taking a very practical, example-oriented approach with lots of working code samples.

BTW, if you're not interested in Go in particular, you could always adapt the code samples to whatever your favorite game is as you work through the book. Could be backgammon, Scrabble, bridge, whatever. I think that would be a great way learn the material!
194758 (5) [Avatar] Offline
#3
Agree. I was a bit perplexed reading chapter 1 as it should be a pre-requisite of the book. Readers who don't know the subject won't be able to get it in a single chapter. It wasn't a big deal as the chapter is short. So far I give the book thumbs up. Hope the author can keep it up when explaining tougher subjects and have good coverage of the subject.

Thank you for writing the book.
maxpumperla (11) [Avatar] Offline
#4
Note that the prerequisites can not fully be specified by the authors alone, but are the product of discussions with the publisher. There were quite a few ideas on what should or shouldn't go into this book.

In the end we settled for a (hopefully) smooth "introduction to deep learning by covering a complete example" with very little prerequisites. Actually it's just high-school maths and feeling comfortable with python (not a total beginner). We want to open up a few core deep learning techniques for a general developer audience.

While that means that towards the end we can't go into every noteworthy detail, readers that started without prior knowledge should have no problem picking up, say, the AlphaZero paper and "get it".