Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[§k]

This question probably sounds stupid because I don't know much about either AI or game dev,

Is there any roguelike where the monster is given a handleful of spells or attacks, and it learns by itself how to kill the player character?

Since AI has played starcraft, surely someone has let it play a roguelike? So a developer makes a roguelike, but leaves out monster AI part. Then if we let AI play it, then this game will finish itself. This sounds wonderful.

[Rolan7]

Angband games support borg-players, but as far as I know they're still merely rules-based.  So are the default AIs for games like Starcraft. 

The relatively new 3rd-party AIs for Starcraft and Age of Empires which used machine learning are pretty cool, but they're specific technology and still guided to a large extent by their developers (from my limited understanding).

You might be interested in this fun video about an AI learning to play some classic NES games.  There's a lot of talk on how he defined "good" behavior to the AI, such as increasing score, which often resulted in short-term strategies which weren't always helpful towards actually winning.

The video ends with Tetris, and a pretty awesome punchline of sorts.  Since the video starts pretty academic, here is a timestamp to 30 seconds before the punchline.  You might want to check that out, then skip around.

[§k]

My prosky is down so I can't watch youtube now. And I read the Roguebasin articles on AI.

As I understand it, making a real AI is more difficult than just designing a pseudo-AI and coding it. That's probably why people still handmake the monster AI in their games.

Now I am trying to make a small roguelike, and that's why I become interested in this topic. The roguebasin articles are really informative and inspiring. I realize that coding monster AI is actually not as difficult as I thought.

[Folly]

All monster decision-making algorithms in all games constitute some degree of artificial intelligence. We don't give monsters human-levels of intelligence because we don't know how to make AI's with human levels of intelligence, it's as simple as that.
In roguelike games the monsters don't generally live long enough for any sort of learning to occur anyway. And if you found another game genre where sophisticated AI would be more effective, I doubt it would be very popular; most people don't enjoy losing to opponents who outthink them at every turn.

[Iduno]

Real AI (machines that can do anything close to what we consider thinking) is pretty far off. Actual machine learning is still pretty crap. Facebook had a copy of the video of the Christchurch shooter, and the "AI" they use to flag videos couldn't learn what a video was well enough to identify other copies.

What you've seen in video games is a pretty decent decision tree, which probably hasn't developed much since Battlechess.

[MagmaMcFry]

Check out Smart Kobold. It's a short roguelike where your aim is to raid a small kobold lair, but the kobolds are actually really intelligent. There's no machine learning involved, but the AI is still miles better than any other roguelike AI I know about.

PS: It's very easy to get completely spoilered for that game. Try to find a way to victory yourself before looking up any strategy guides. (Since the kobolds are smart, you need to be smart too.)

[Cathar]

The relatively new 3rd-party AIs for Starcraft and Age of Empires which used machine learning are pretty cool, but they're specific technology and still guided to a large extent by their developers (from my limited understanding).

Well actually no. Alphastar uses agents trained originally from human datasets, but that's only the core of the AI. Stuff like moving units, building a base or launching an attack. The rest of their training is made by launching some darwinian tournaments between different agents with the victor of each pool fathering more children while losing agents would disappear. They can't be really guided by they programmers because their programmers...don't know how to play the game.

On launch, the top 5 alphastar agents absolutely obliterated the best players of the western hemisphere. Their programmers explained those agents had about multiple centuries worth of accumulated experience.

But yeah, they are specific technology ; they can only play Starcraft II, Protoss v Protoss on one specific map. And it is cutting edge technology. Don't expect indie dev making stuff like that

[§k]

Check out Smart Kobold.

This game looks very interesting! I'm downloading it right now. I heard it before but somehow didn't think of checking it out.

And it is cutting edge technology. Don't expect indie dev making stuff like that

What would be the cost of letting a monster machine-learn some basic roguelike tricks from scratch, say, kiting, pillar dancing, retreating to corridor when being surrounded?

[§k]

Check out Smart Kobold.

This Smart Kobold game is great fun. You basically play a Giant trying to raid a player-run fortress(Not the adamantine axelord type, but the trap-and-exploit type). You'll see trap corridor, civilian alert, crossbow militia... And the kobolds have even better composure than our ordinary dwarves.

[Iduno]

Check out Smart Kobold.

This Smart Kobold game is great fun. You basically play a Giant trying to raid a player-run fortress(Not the adamantine axelord type, but the trap-and-exploit type). You'll see trap corridor, civilian alert, crossbow militia... And the kobolds have even better composure than our ordinary dwarves.

I love it when they

Spoiler (click to show/hide)

break their bows when cornered.

[McTraveller]

What would be the cost of letting a monster machine-learn some basic roguelike tricks from scratch, say, kiting, pillar dancing, retreating to corridor when being surrounded?

It's simpler to pre-code some "primitives" and then only rely on the machine learning for how to string the primitives together.  If I recall, this is how Boston Dynamics does their scary robots - they don't have the robot learn the basic movements, but only how to combine the pre-programmed movements into more complex behaviors.

So you could very easily build the basic verbs, and then just have the machine learning figure out the best sequences to combine those verbs.

[§k]

It's simpler to pre-code some "primitives" and then only rely on the machine learning for how to string the primitives together.  If I recall, this is how Boston Dynamics does their scary robots - they don't have the robot learn the basic movements, but only how to combine the pre-programmed movements into more complex behaviors.

So you could very easily build the basic verbs, and then just have the machine learning figure out the best sequences to combine those verbs.

That sounds totally doable! If no roguelike has tried this, I'll probably try it out.

[Starver]

When the monsters in a dungeon go "Oh, you again..! And as a rogue, this time, I see. Look, before we start there's a healing potion over there to the left, but watch out for the trap. Shall we play Rock Paper Scissors for who gets to pick up that helm?" then it may have gone too far. Especially when they intentionally lose the game for the helm because they know it's cursed.

[§k]

When the monsters in a dungeon go "Oh, you again..! And as a rogue, this time, I see. Look, before we start there's a healing potion over there to the left, but watch out for the trap. Shall we play Rock Paper Scissors for who gets to pick up that helm?" then it may have gone too far. Especially when they intentionally lose the game for the helm because they know it's cursed.

How does one intentionally lose a turn based rock paper scissor?

If I can let the AI make something half smart as a real life cockroach in a roguelike dungeon, I'd count it success.

edit: If I make a roach-wit monster without using AI, I'd count it success too. I'm not sure if AI is really an efficient method.

[Starver]

How does one intentionally lose a turn based rock paper scissor?

Ah, now, that's how clever they'd be!

(As far as being as smart as a cockroach, start with your basic attract/repel instincts, augment on some learning behaviour for novel (not previously defined/definable) situations, a basic situational awareness and memory capable of exploring the map. Set 'cockroach' against 'cockroach', with select 'roaches given player-like abilities to use and exhibit, set it all off in unattended mode and see if there' s any emergence of simple differentiated tactics vs Tank/Mage/Projectile-thrower entities that might be transferable to use against human-played opponents, maybe?)