AI

[radinms]

anyone create the AI that can do 100 streaks by using "deep learning"?

[Jarlyk]

I've thought about trying to create an AI for crawl as a way to experiment with various 'learning' techniques. Honestly, though, I'm not sure if much 'learning' is necessary. Most of the things that players learn intuitively (in particular, which monsters are dangerous under which situations) can be estimated more directly by a bot, since the source code is provided and hence almost full information is available. The biggest asset for a bot in crawl, assuming you're not trying to minimize turn count, is that the bot has unlimited patience. You can use the most degenerate tactics possible to minimize risk and you can run probability simulations directly for every encounter, taking a lot of the intuitive guesswork out of the 'fight or flight?' decision. You can explore as conservatively as possible (minimizing exposure to unexplored territory while preserving an escape route) and almost always lure back to safe territory before fighting.

I'm not saying that a bot with the above level of tactical/strategic reasoning would be easy to implement, but none of the above really requires 'learning'. It's more a matter of numbers and direct probabilities, then minimizing risk in tactical encounters. There aren't many situations in crawl where locally optimal tactical decisions have a strategic detriment, so you can largely perform local optimization. The main ones I can think of are the OOD timer and potentially 'hunger', but these are not particularly tight constraints. This situation is very different from games like Chess or Go, where subtle early misplays can have profound long-term implications for your overall strategy*.

It would be interesting to see what areas could be addressed with a 'deep learning' framework, though. Most likely this would be applied to the strategic layer (responding to available items and planning overall skill progression), since these decisions are longer-reaching in their effects. The trouble, though, is that these decisions may not actually be all that important.

*Note that I don't think this is a bad thing; crawl is way too long of a game for it to be fun for a subtle early mistake to result in a later inevitable loss.

[le_nerd]

I have given thought to it, since the only input the deepmind software need is a score that increases with progress, and 2d pixels. Should be a cakewalk! Have not yet tried, and am looking forward to see the really degenerate optimizing stuff the trained AI will do to increase score (breadswinging!).