First NetHack bot ascension
reddit.com
reddit.com
telnet nethack.alt.org
If you want to start playing, telnet there, create an account, hit "p" to play a game and then "?" to read basic help. http://nethack.alt.org/ will keep logs of your game as well as other stats.
This post also mentions previous bot attempts such as the Tactical Amulet Extraction Bot (TAEB), which is also worth looking at: http://taeb.github.io/
Warning, SPOILERS: If you're okay with spoiling the game to some extent (e.g., solutions to common puzzles, which corpses are safe to eat, strategies), check out the NetHackWiki: http://nethackwiki.com/wiki/Main_Page
It is definitely more fun to try playing the game for several hours without reading the spoilers, but you're not going to figure out stuff like the Invocation Ritual without reading the spoilers.
https://groups.google.com/forum/#!topic/rec.games.roguelike....
This did involve use of wizard mode (not for the unspoiled ascension itself, but beforehand for learning stuff).
People often ask how someone would learn how to perform the Invocation without spoilers. One intended answer by the DevTeam is to pay the Oracle for a series of major consultations, one of which (somewhat elliptically) explains the Invocation. The Oracle consultations aren't considered spoilers if you pay for them in-game.
What's you're not going to figure out spoiler-free are things like magic cancellation. Which is maybe not literally essential to winning, but you're going to have a tough time without it.
Of course, you'd have to be good enough at the game to survive to the Oracle and have the cash to pay for the major consultations. And you'd have to know that information source exists.
Since Nethack is a game where not knowing the details of some minor edge case can and will kill you dead, or deader than dead, I don't know of anyone who actually relied on that (though there may very well be someone).
It's really, really neat. Even to those of us who know little or nothing about the game, from the nested Englishy descriptors piled up into short conditions for things-the-bot-might-want-to-do, the basic strategies can be discerned...
(def desired-shirt
(ordered-set "T-shirt" "Hawaiian shirt"))Absolutely cannot tell if you're fucking with me, or if this is an actual game mechanic - and if so, why this shirt would be desirable.
Welcome to the world of nethack! The amount of special case situations that have been programmed into this game over the last few decades are truly amazing. (Mostly because it somehow still works.)
I still remember the first time I ran into a kitchen sink.
http://tvtropes.org/pmwiki/pmwiki.php/TheDevTeamThinksOfEver...
> Additionally, throwing a cockatrice corpse up (<) will result in it hitting you on the head and, unless you're wearing a helmet, petrifying you instantly. Cause of death? "Petrified By Elementary Physics"
Some of these I knew, most I did not. Thanks for sharing that.
The Quantum Mechanic monster is basically one big physics joke.
Its attack will randomly teleport you, and eating a Quantum
Mechanic corpse will give you the intrinsic speed ability (or
remove intrinsic speed from you if you were already fast). Upon
death, the Quantum Mechanic also has a small chance to drop a
box. When you open the box, a cat named Schrodingers Cat will be
found inside. 50% of the time, the cat will be alive, and 50% of
the time, the cat will be dead.
What brings this into The Dev Team Thinks of Everything territory
is that if you look at the game's source code, you discover that
unlike every other box in the game, the state of the box's
contents (whether the cat is alive or dead) is not determined
until you open the box and look. (or (if (and (weak? player)
Or this probably simple function(?):
(defn- want-protection? [game]
(and (< (:protection (:player game)) 3) (> 15 (:xplvl (:player game)))))
def want-protection(game) =
game.player.protection < 3 &&
game.player.xplvl < 15The second function is easier for me to understand: it says that the player wants protection if the player has less than 3 intrinsic protection and has an experience level higher than 15. (In NetHack, gods grant intrinsic protection under some circumstances, which causes one to become better armored even without wearing additional physical armor. It's basically like a kind of magical armor on one's body that's independent of any physical object.) The idea seems to be that a player with at least experience level 15 ought to have 3 or more points of protection, and if the player doesn't, the player should try to go and get them.
If you're not used to reading parenthesized prefix notation, you might have a look at
https://en.wikipedia.org/wiki/Polish_notation
https://en.wikipedia.org/wiki/S-expression
https://en.wikipedia.org/wiki/Lisp_%28programming_language%2...
[1] In another example of NetHack's complexity, it's also possible in some circumstances to eat objects other than food, by turning into a monster that's able to eat the objects in question.
The first one is missing a whole bunch of context. But here's a start: The first thing inside parentheses is the operator/function, the rest are arguments/operands. This one is best viewed from the inside out, so... a method ending in a question mark is most likely a predicate, i.e. a true/false test returning a boolean. So we're testing for a weak player. Outside from there is an 'and', so we're looking to combine that test with another one (further to the right, that you didn't show). Maybe checking if the player is both weak and hungry, as that may mean he'll croak soon.
Every paren pair ('S-expression') in Lisp (err, Clojure) returns a value, so the "if" is not so much a control-flow construct as it's like C's ternary op: [ cond ? iftrue : iffalse ]. Its first argument will be a test, like that weak-hunger thing. Then you will have one or maybe two other arguments. The first is returned if the condition is true, the second one if false. There may not be a second argument, in which case the return value will be 'nil' or maybe 'false' - they're sorta equivalent anyway.
Finally, the "or" will evaluate and return its first subexpression if it's true, or also evaluate the second one if not and return that. You can actually have as many arguments to an "or" as you like - they will be short-circuit-evaluated until one of the arguments is true, or if you hit the right paren without finding a true, you'll get a 'false' ('nil' ?).
------------------------------------
The second example is a function definition. By convention, 'want-protection?' will be the name of a predicate, i.e. something that tests and returns true/false.
'game' is the single argument to that function. From usage later on, it looks to be a map, or you might think of it as an associative array, or a key-value mapping.
The first half tests whether the 'protection' value of the player, which itself is the 'player' value of the game (in C: game.player.protection, except those components are keys rather than fields) is less than (see the '<' operator on the left?) 3. The second half tests whether (I'm guessing) his experience level is greater than 15. The 'and' combines those, i.e. if he's at least a lvl 16 player and his prot value sucks at less than 3, he'll want protection (so return 'true').
I read the reddit thread, and while it was in English, it made not a lick of sense to me. Now I know how non-engineers feel sometimes. :)
BTW, can someone tell me why this is such an amazing accomplishment? I know Nethack is very old, so is this a case of a complex problem space or just no one has tried before?
Combat in non-enclosed spaces is hard do to ranged attacks line-of-sight, etc.
I'm slightly surprised it has taken this long to do with pudding-farming, as that lets you get plenty of really powerful items (as well as other benefits from sacrificing the corpses)[1]. There are a few parts of the game that are super hard for a bot even with the equipment of pudding farming, but I would have thought a few thousand runs would be enough to luck through them with simpler rules.
1: Puddings are creatures which can split in two when you hit them, and each new pudding will count as a corpse (which you can sacrifice to your deity for benefits) as well as having items. I don't recall all the details, but some of the rare drops from black puddings are really powerful. The idea is you take a really crummy weapon, so you don't accidentally kill the pudding and then hit it over and over again to get lots of treasure. In this run it looks like the bot used a skeleton key as a weapon.
The main deterrent for a non-bot player doing it is that it is really boring (There's a quote somewhere from the devs that is roughly "we implemented a penalty for pudding farming, it's called 'pudding farming'" Obviously the other deterrent is the prestige of beating the game without pudding farming, and that does apply to bots too.
That makes me a lot more optimistic that the bot is close to being able to ascend without farming, especially since I believe its character already had multiple amulets of life saving and a bag of holding even before starting to farm.
The fact that someone wrote a bot to think on its feet enough to do this, is astonishing.
You should try it out in your free time, check out alt.org, there's a server where you can play and the irc channel #nethack on freenode (I believe) is very helpful and there's a few wikis available.
But beware, it becomes very addictive. The 'one more turn' rule applies heavily here.
I agree that the bot ascension is an astonishing achievement, and one I didn't expect to see for many years!
Crowning moment of awesome entry.
The sigil-based graphics are part of what make the complexity possible. And - the dev team have coded in all manner of combinations of things, many of which get you killed in plausible/amusing/annoying/complicated ways.
There's a culture of players looking at the source code while they play. The source is really antiquated C.
If you decide to have a go, I'd recommend printing out a copy of the manual, avoiding the spoilers for the first couple of dozen games at least. Just try to learn your way around the dungeon just from that manual, and experience the complexity and sense of mystery.
It's not a particularly 'good' game - grindy. A vogue at the moment in roguelikes is for games with as little complexity as possible. For example, I've spent far more time studying the C code to this game than I have playing it: http://trogrd.tumblr.com/post/61949503427/zoo
I don't think that is relevant, I have played the GUI version for many years, and it's perfectly usable.
A letter 'c' is a cockatrice or a pyrolisk; those are both quite dangerous so I suggest you learn to recognize them for your own good.
You get potions, scrolls, weapons etc but have no clue what they are until you use them (some scrolls will help you identify what other things are in advance). Nothing about them will give you any clue (scrolls have names like "AFEWA AWEBRPO", potions are just colours etc.) To survive you need to use them though. You need to drink, but the potion might be one of weakness, poison, blindness, confusion etc. You need to eat regularly, and food is scarce, but it also might poison you. A ring of slow digestion is helpful there, but until you use an identify scroll on a ring you won't know what it is. It could end up being a ring of attract monster, for example.
• Many options are available:
- Many different actions available at any given time (move, attack with a weapon, Zap using a wand, Quaff a potion, Wear a different piece of armor or equipment, Wield a different weapon, Read a scroll, Spellcasting, etc)
- Many different options available within each action (place to move, spot to place a fireball or lightning bolt, place to teleport to (!!), which piece of armor or equipment to use)
As a very rough estimate from someone who hasn't played NetHack much, I'd guess that a mid-level character might have 10^4 or so options immediately available to them at any given time, though many of these are combinatorial results of having thirty different spells that can be targeted on any of the hundred or so visible tiles. The branching factor for Chess is about 10^1.5, and Go is around 10^2.
• Options are difficult to evaluate:
- On an immediate mechanical level: If you want to use teleport, what's the best place to teleport to? If you want to cast Fireball, how do you use it as effectively as possible? If you want to shoot into a group of enemies, where do you aim?
- On a tactical level: Would this monster be more effectively dealt with by spending some energy to cast Fireball, using a charge from your Wand of Magic Missiles, or attacking it in melee and risking taking some damage?
- On a strategic level: Would it be more useful to clean out the rest of this level or immediately go down the stairs? What set of armor and weapons would result in the best chances in melee against the enemies I expect to encounter in the near future? Which of my items should I cast Polymorph on to maximize the chance of receiving a random item that would be useful to me?
- On a mathematical level, NetHack suffers significantly from the horizon problem. That is, many immediate tactical or mechanical decisions have strategic consequences that require an AI to simultaneously consider all of the above problems. You can't just have a mechanical planner, a tactical planner, and a strategic planner; you have to have a unified AI that can realize that it'd be a good idea to not kill that one enemy in the corner because it can kill it and eat it two hours later on its way back up to the surface.
This is an even nastier issue. Even humans can't do this effectively for NetHack. For example, if you look on the NetHack subreddit, one of the current top posts is "Help me make it out of this group of enemies alive!", and everybody who comments has a different idea of what the best solution is.
My neutral human wizards always have either Frostbane or Grayswandir (Magicbane as backup/for MR), robes, silver dragon scale mail, boots of speed, enchanted T-shirt, gloves of dex, etc.
Those swords barely do any more damage than Mbane even in the hands of combat classes.
The essence of roguelikes is that of Diablo and other similar games: you're walking around killing things, picking up items and advancing levels. In most roguelikes, however, your freedom of movement, details of the environment, number of choices to make are on a whole another level. Why have 3 (or 5) PC classes, when you can have 20 or 22, each really unique in strengths and weaknesses, utilizing different tactics and strategies? And PC race matters too, so the number of available combinations is ~(12 times 22) before you even started playing!
One in-game example of how detailed and complex roguelikes tend to be: some levels are flooded and you can't get past water without knowing how to swim... Wait, no. You of course can go into the water, no matter the swimming skill. Without swimming skill - and with swimming not trained very high - you'll lose some HP due to drowning. You can reduce the chance of drowning by leaving your heavy equipment on the ground. On the other hand, going into the water in this shiny metal armor will make it rust. After a while. Note that there are metals that do not rust, these are safe. Also, the scrolls and magic books, if not wrapped in a waterproof blanket, become unusable after some time in the water. You also obviously don't want to try lightning based magic when standing in the water. But hey, you can just build a bridge! If you have a skill for it, a hatchet and some trees nearby. Or, and this is my favourite way, you can just freeze the water with (most of) ice based spells. But don't do this when being chased by a fire using enemy: your ice bridge can easily melt. Also, keep an eye on how much your equipment weights: bridges made of ice are not very solid. You could also get a pickaxe and try to dig your way around the water, but this takes time, so maybe a wand of digging would be better...
And that's the number of options available in only one of hundreds of different and interesting situation the game puts you in. It's an experience as close to classic table-top RPG as possible without a human as a dungeon master. It also makes you forever unimpressed with "interactive environments" and "complicated plots" of modern games.
Dwarf Fortress ruined modern gaming for me. I keep thinking "Why do they keep making new games when they could just make the old ones bigger and better?"
People love it when their games evolve over time. A new, small feature every week will keep people feeling refreshed and invested in the game.
Various people have used (a version of) Super Mario Bros. as a demonstration of planning AI recently. As a result, we see that people can write planning algorithms, based on imperfect information about the game state, that can complete a straightforward Mario level with some competence, but not up to the level of competence of, say, a typical '80s child.
Ascending in Nethack is an insanely difficult accomplishment that few people ever achieve. The game is much more complex than a Mario level, though perhaps the non-real-time aspect of it allows for more sophisticated planning.
There's a bot with several wins in Dungeon Crawl Stone Soup, another roguelike. The combinations it has won are pretty rote, but it's still very impressive, particularly since I don't think it's even capable of knowing that much about the game (it uses Lua handles to pick up information about the state and I believe there's a fair chunk of the game that isn't exposed that way.)
so, the bot is also a better player than me.
Also, the bot can safely walk along the edge of lava without falling in due to a typo. :-)
With the more popular games it reaches a point where they are barely even playing the game anymore, instead running on auto-pilot and hoping to make the least mistakes and get the most favorable RNG with minor things like enemy AI patterns.
Speedrunning a game with roguelike randomness is a completely different beast, requiring actual mastery of the game as it was intended to be played.
I agree here, but I just want to emphasize that Any% speedruns (runs where you purposely exploit glitches or flaws with the game engine to trigger the final cutscene as fast as possible) still take incredible amounts of skill. More skill than playing the game as it was intended to be played, in a lot of cases. The current Ocarina of Time route involves exploiting several glitches that require pressing the correct button combination on a specific frame (N64 games run at 20fps), jumping at just the pixel perfect angle to slide through a model's corner/collision tracking, etc. Much more difficult stuff than beating the game straight through. Super Monkey Ball 2 (Gamecube game) speedruns require pulling off maneuvers trickier than anything the levels were explicitly designed for, the difficulty of which is hard to overstate. That game is hard. Speedrunning a game with true RNG everywhere like Nethack is definitely a beast of its own since you can't memorize routes, but even runs with completely static routes are incredibly difficult to pull off.
Such an ascension is aptly named "Zen Ascension".
For those of you who are looking to get into nethack, or play it all the time... (shameless plug) I've been working on a project for some years that helps ease some of the monotony of playing.
https://github.com/samfoo/noobhack
For example, it keeps track of where you see shops, so you can come back, maps out levels and what you've seen on them, and auto-price identifies things you pick up.
There's a rudimentary plugin system so you can write your own extensions (more coming soon).
The author's list of milestones illustrates the key components of this achievement, as well as some limitations:
https://github.com/krajj7/BotHack#milestones-reached
Awesome work, Jan Krajicek!
If anybody gets introduced to Rogue-like UI for the first time because of this article, and likes it, I also recommend checking out Dwarf Fortress. Similar but very very different in ways that people tend to either love or hate.