5 ms·
Bandit Algorithms Book [pdf]
- clickok 8y agoI skimmed through this and have already found a bunch of interesting sections, but there's also a ton of background information on topics related to bandit algorithms. The authors say that this is the first draft of the book submitted to the publisher, so I suppose it's nearly complete? More details available at the site they put up, http://banditalgs.com/ http://banditalgs.com/
- shoo 8y agoReaders who enjoy banditry may also enjoy John Langford's http://hunch.net http://hunch.net
- daleroberts 8y agoCool, nice to see that Tor was a student at ANU.
- HugoDaniel 8y agoIs this the book that is going to make me a poker master player ?
- srean 8y agoIf you play long enough it will make you regret less
- dsvmn 8y agoI really appreciate sharing the book. However, to everyone in charge with naming these files, please don't call it "book.pdf". It makes everyone go to their computer and rename the file after downloading it so that they can find it later. Give it a more intuitive name. Thanks
- sureaboutthis 8y agoWell that's really great! What is it?
- tomkat0789 8y agoNever heard of bandit algorithms before! Or if I did I didn't recognize it as something different from probability. What have people around here used them for?
- zdkl 8y agoThis rust project, to manage the number of threads in a monero miner afair. https://github.com/Ragnaroek/mithril https://github.com/Ragnaroek/mithril
- ur-whale 8y agoDoesn't alphago use some form of Bandit algorithm in their MonteCarlo code?
- magoghm 8y agoI believe that Monte Carlo Tree Search, used in AlphaGo, does work using bandit algorithms. On top of that AlphaGo uses Reinforcement Learning, which also uses bandit algorithms (in Sutton & Barto's book, "Reinforcement Learning: An Introduction", all of chapter 2 is about multi-armed bandits).
- haffi112 8y agoYou can use it when determining the best solution being tested in as few trials as possible. Say you are selling a product and you are AB testing something related to buying the product. When a user visits the site you ideally want to give him the version you are more confident is better. By using a bandit approach you can determine if say option A is currently better (w.r.t. some confidence bounds). After each visit you can update the bounds and after sufficiently many visits you have a winner. The main difference to more traditional AB testing is that the process is more adaptive and less time is wasted on exposing an inferior product to the user.
- bochi 8y agoBandits are probably one of the most underrated machine learning algorithms. One possible application is recommendation systems. Shameless self promotion. I wrote an article about it: https://towardsdatascience.com/how-not-to-sort-by-popularity-92745397a7ae https://towardsdatascience.com/how-not-to-sort-by-popularity...
- pronoiac 8y agoThis came up a couple of days ago: https://news.ycombinator.com/item?id=17637683 https://news.ycombinator.com/item?id=17637683
- joshuamorton 8y agoIt always makes me sad that Thompson Sampling isn't (or at least doesn't appear to be) mentioned alongside things like UCB1. Its theoretically optimal, and relatively easy to grok, and not significantly more difficult to implement.