3 ms·
one thing i started doing in recent years is just borrowing heaps of books from the library. trawl the catalogue, see what's ready for borrowing, borrow half a
by shoo 8y ago
one thing i started doing in recent years is just borrowing heaps of books from the library. trawl the catalogue, see what's ready for borrowing, borrow half a dozen ones that look interesting or are on the reading backlog. then start reading a few of them in parallel. if some of them don't hold your attention, stop reading and focus on the ones that are interesting. take them all back to the library when they're due, read or not.
edit:
okay, so now for the obvious question, what's the correct way of framing the which-books-should-be-finished problem, and what's the optimal strategy?
is this a bandit algorithm thing? at each interval t, sample a page from a book according to gittins indices to minimise lifetime regret? does it need monte carlo tree search?
- rusk 8y agoMassively Parallel Speculative Evaluation. I like it!
- noelwelsh 8y agoNot a classical bandit because the arms have a finite lifetime (the arm is no longer active once you read the whole book) and you know this lifetime. Definitely can be viewed as some form of RL problem. For me the challenge is how you define rewards and hence what information reading part of a book gives you about the total reward for completing the book. Maybe assume some smoothness criteria, so that the reward for reading part n of a book is highly informative of reading part n+1, but less informative of part n+2, and so on. More formally, each book gives some total reward on completion and that reward is distributed according to some stochastic process. Greedy strategies probably work reasonably well.