https://github.com/krajj7/BotHack

https://www.reddit.com/r/nethack/comments/2tluxv/yaap_fullau...

Edit: I remember looking at BotHack at the time and being surprised at how extensively and elaborately hard-coded the strategy and evaluations were. This is definitely not machine learning!

Edit 2: HN discussion at the time: https://news.ycombinator.com/item?id=8990869

Here's an example method: https://github.com/krajj7/BotHack/blob/master/src/bothack/be...

    (defn- read-book [game]
      {:pre [(have game book #{:noncursed})]}
      (if-not (invocation-complete? game)
        (if-let [[slot _] (have game "Book of the Dead")]
          (with-reason "reading the book"
            (->Read slot)))))

There's a long history of using bots (they weren't really smart enough to be called AI, but they did "learn") to play games like this.

Long before Nethack there was Rogue. And then, 37 years ago, there was Rog-O-Matic. As Wiki describes it:

Rog-O-Matic is a bot developed in 1981 to play and win the computer game Rogue ...

Described as a "belligerent expert system", Rog-O-Matic performs well when tested against expert Rogue players, even winning the game.

https://en.wikipedia.org/wiki/Rog-O-Matic