It was very carefully written in Clojure, and the result is some beautiful and fairly self explanatory code. The bit that ties everything together to define the way the bot behaves:
Looks like this repository is mostly concerned with standardizing the API to make it easier for others to build agents. It did come with code for an example agent, not sure how advanced that one is.
The attached example_run.gif ended with "Agent76001 killed by a gnome lord, 2019".
I think the maximum score we've sene so far is maybe something like 4,000 or 5,000. They descend down to dungeon level on average five or six
sometimes luckily, going down even dungeon level 15
I’ve been playing for maybe 100 hours over the course of a decade, and have reached level 8 maybe once or twice. I just keep getting killed by some stupid ant or wolf or whatnot.
I've been playing on and off for 3 decades. Finally started winning when I used the iPhone and Android versions.
Something about the UIs makes me better than that PC interface.
A large part of attraction of NetHack Learning Environment is availability of large human play dataset, including expert demonstrations. While I agree playing dungeon master sounds fun, that game lacks such data.
For one thing, you could make the combat more interesting and varied. The monster AI is pretty simple and predictable.
Soldiers or orcs could communicate and coordinate, for instance. So instead of simply walking the shortest path to the player to be dispatched one at a time, they could harry the player, use ranged weapons, retreat strategically, cut off the player's retreat, etc.
You could have something similar for tame monsters.
It would change the gameplay significantly, and require careful consideration of game balance. I think it would be more of a whole new variant, rather than something to tack onto an existing version of the game.
AI driven NPC interaction could also provide some amusement and make the game less predictable. You could give shopkeepers and other NPCs distinct personalities, where there behavior is driven by complex motivations.And I'm sure lots of other stuff.
Edit: a little digging and "A tilde is a multi-user online system designed to be a "home" for users by providing services such as shell access via SSH, email, small site hosting (on www, gopher, gemini), chat, or other programs. It is a loose term, akin to a public access unix systems (a.k.a. pubnixes), but with some connotations of a small community" https://tilde.wiki/wiki/Tilde
https://www.reddit.com/r/nethack/comments/2tluxv/yaap_fullau...
Previously discussed: https://news.ycombinator.com/item?id=8990869
The code: https://github.com/krajj7/BotHack
It was very carefully written in Clojure, and the result is some beautiful and fairly self explanatory code. The bit that ties everything together to define the way the bot behaves:
https://github.com/krajj7/BotHack/blob/master/src/bothack/bo...