If you're interested in these types of algorithms, I'm writing a book / running an online course. You can sign up and read past content here: http://bandits.myn

If you're interested in these types of algorithms, I'm writing a book / running an online course. You can sign up and read past content here: http://bandits.mynaweb.com/ http://bandits.mynaweb.com/ Next section is coming very soon. Just gotta finish the last visualisation. (Interesting that both bandit algorithms and Angular + d3 feature on HN today.) I haven't had the time to read either article closely (at a conference) but the main issue I'd be concerned with is non-stationary rewards. What is interesting today is not interesting tomorrow. A simple hack to get around this is to decay the prior over time. Other issues to consider include: - the amount of data available. Even HN might not have enough readers to get sufficient data on each arm. Here you might use a similarity measure between arms (articles), such as tf-idf so votes for one article can inform the ranking of other articles. - the effect of ranking on votes. The highest ranked articles probably gets more votes just by virtue of being at the top. There is some work on submodular bandits that might be relevant here, but this research is very new.