Monster AI (was: Dungeon Layout)

[galehar]

I don't like making monster AI smarter. Mostly it seems like it would make the game less tactically interesting - if you can't predict monster actions then you can't factor future monster actions into your tactics.

So you're saying that if monsters keep being dumb and predictable for the whole game, it's tactically more interesting? Using the same cheesy tactics from D:1 to Zot is more interesting than facing unpredictable and smart opponents that you have to adapt to? Is playing an RTS against a dumb AI also more tactically interesting than playing against a smart one or a human?

Also, in many cases it would increase the role of spoilers (look at ranged AI for an example of this happening).

What makes the game spoilery? It's when something is illogical or unintuitive so it has to be learned with experiments, spoilers or source diving instead of being logically deduced. Prior to the ranged AI, the duck behind a corner tactic was spoilery, because it was illogical. The new ranged AI is less spoilery, because it's intuitive. A new player would expect the monsters to behave like that, because that's what he would do. The spoileriness that remains comes from the flaws of the AI, like not avoiding monsters and plants to get LOF. This is illogical, thus spoilery.

When I fix that, I'll add some randomisation to it factoring things like how intelligent the monster is, how much it likes ranged over melee and is it part of a band with a living leader. I think the end result will be logical and intuitive behaviour (so not spoilery). I also hope it will make encounters with high level monster bands more challenging and interesting. I really think adding more varied AI and better monster bands (complementary abilities) is a better way to increase the mid-game difficulty than adding tougher dudes.

[mumra]

It's very easy to learn exactly how monsters move from a few minutes of play - normally they take the shortest path to the player. This is an intuitive way for them to behave because the game is a game - nobody is going to expect it to approximate reality, and even if they do, they'll learn how things actually do work very quickly because monster behaviour is simple and simple things are easy to learn. It is indeed pretty spoilery for monsters to move away from a corner by 3 or sometimes 4 squares if you move around a corner oh and if they can't do that they'll follow you normally oh and if you happen to want to keep them from attacking you for a bunch of turns and have a bunch of pillars you can do that but only by moving in a really complicated pattern that nobody would get right on the first try.

A spoiler is defined as "something the player gets an advantage by knowing but cannot work out just by playing the game". AI manoeuvres simply don't fit this - no matter how complex the AI, a human is still smarter than a computer, and can actually spot the pattern pretty quickly. This then becomes tactically interesting because you have to work out the best methods for your character to outmanoeuvre that AI. There's no book of spoilers for that one because the "optimal" approach is entirely dependent on character, playstyle and environment. Something being complex to work out (in other words, a puzzle) is not the same thing as "a communicable piece of knowledge that gives you an advantage that you could only otherwise discover through luck".

And ... running around a pillar is complicated? Really?

[mumra]

If you can actually work out what ranged AI does just by playing then that's pretty impressive because I haven't been able to.

I couldn't say I know exactly what it's going to do at any given time but, y'know, that's the thing about Crawl, things are randomised so you can't always predict them ... But in Crawl as in any other game I gradually get a feel for what the AI is likely to do in certain situations. And in this case that's the best you can do, because the AI decisions come down to the RNG; and you couldn't gain an edge even by looking at the source, so it's not spoiler territory, which is all I was saying.

[njvack]

"Better" AI doesn't mean "more like a human" AI. If the computer played like a human, Crawl would be impossible, I think -- orc priests would smite every turn, because that's their best move, always. Wizards would turn invisible, throw flame, walk away or blink if you got near. Centaurs would kite you.

"Better" AI may, however, mean more interesting, less-predictable, more-dangerous behavior.

Monster behavior that's both neither braindead nor deadly is inherently really hard in games that are as asymmetric as Crawl.

[dtsund]

There are upsides to predictable AI an an against-the-computer, asymmetric game like this. If the computer's behavior can be predicted, this allows the player to reasonably calculate the game state several turns in advance, and also makes it reasonable for there to be situations where doing so is required.

[galehar]

And what tactics do you consider "cheesy," anyway?

Corridors? Wait, I know your answer. If corridors are bad, remove them. However, I'd rather ruin them and give some monsters enough AI so they can flank you. Seems like it would be tactically more interesting than open levels with dumb monsters.

It's very easy to learn exactly how monsters move from a few minutes of play - normally they take the shortest path to the player. This is an intuitive way for them to behave because the game is a game - nobody is going to expect it to approximate reality

I think we have a different vision for what crawl is. I am under the impression that you see it as a chess-like game and monsters as bunch of stats and special properties with simplistic behaviour. I see crawl as a game of dungeon exploration in which you kill monsters.

I guess we just disagree there. But I'd like to point out: the early game manages to be very interesting without complicated AI.

It's true, and I have no intention of changing it. However, do we really need to have more and more of the same thing? We're always trying to differentiate monsters why not have variations in how they behave?

"Better" AI may, however, mean more interesting, less-predictable, more-dangerous behavior.

Indeed. When I say I want to improve the AI, I'm not saying I want it to behave optimally. I want more variety and less predictability.

There are upsides to predictable AI an an against-the-computer, asymmetric game like this. If the computer's behavior can be predicted, this allows the player to reasonably calculate the game state several turns in advance, and also makes it reasonable for there to be situations where doing so is required.

There's already a lot of randomisation in the monster behaviour. It's only their movement pattern which is really simple and predictable, but you never know if the monster will shoot, cast, buff, zap or move. Only melee dudes are very predictable since they only move or attack.

[njvack]

I think we have a different vision for what crawl is. I am under the impression that you see it as a chess-like game and monsters as bunch of stats and special properties with simplistic behaviour. I see crawl as a game of dungeon exploration in which you kill monsters.

Ah: is Crawl more like an RPG or more like a strategy game?

I definitely enjoy thinking of it as an RPG. I'm only able to win by thinking of it as a strategy game and thinking of the stats, properties and behaviors.

As I also enjoy winning, these things are actually somewhat at odds for me.

[galehar]

Ah: is Crawl more like an RPG or more like a strategy game?

The strategy is of course much more important than the "roleplaying". And the tactics even more. But I like that this tactical game takes place in a world, have theme and uniques with personality. It's not an abstract game.

[Blade]

Because it's more strategy/tactics than role-playing, though, the mechanics should be conducive to that style. Monsters moving in intelligent ways doesn't add to the flavour; rather, it adds an unnecessary level of complexity.

[mumra]

People are reading too much into "AI". It doesn't mean the monsters will actually be made intelligent. It just means they are given more complex and interesting behaviours that make it harder (and therefore more tactically interesting) to apply simple rinse-and-repeat measures to deal with them. This is clearly only a good thing.

[njvack]

Yup. If we're talking about things like the new ranged AI and (intelligent) monsters not running headlong into dangerous clouds, I think they make the game more fun. If a pack of skeletal warriors flanked you to close off your escape instead of piling up in a corridor, that'd be fun, too, I suspect.

[nicolae]

I like the idea of better monster AI. It would provide a new way to make things more challenging besides "resists everything and does irresistible damage".

[naughty]

People are reading too much into "AI". It doesn't mean the monsters will actually be made intelligent. It just means they are given more complex and interesting behaviours that make it harder (and therefore more tactically interesting) to apply simple rinse-and-repeat measures to deal with them. This is clearly only a good thing.

If it helps we call this 'idiosyncratic AI' where I work, not sure if there's a general game design term for it. You see it a lot in Japanese games where the individual AI of enemies is quite simple but very distinct, you then mix up the enemies to get the tactical play out of it.

From a roguelike perspective I think Shiren and especially Brogue have some great enemies that don't have direct analogues in crawl, e.g.

• In Shiren when you kill and angry soldier it's ghost appears after a few turns and hunts for the nearest enemy, once it reaches it levels the other monster up.
• Again in Shiren there's a tiger enemy that throws other nearby enemies at you.
• Brogue has annoying, thieving monkies kind of like Maurice without the invis.
• Bloats that explode into caustic gas clouds when killed. There's also a version that causes a temporary holes in the floor.
• Goblins, ogres and trolls dancing around destructible totems that buff them.

[XuaXua]

I don't like making monster AI smarter. Mostly it seems like it would make the game less tactically interesting - if you can't predict monster actions then you can't factor future monster actions into your tactics.

This isn't Desktop Dungeons.

The only thing you are guaranteed to find is a loaf of bread on the first floor and eventually an altar for every deity.

[XuaXua]

If you can actually work out what ranged AI does just by playing then that's pretty impressive because I haven't been able to.

If you go out of line of sight, the AI will pursue you. If it is possible to not be adjacent to the mouth of a corridor you entered, it will not be adjacent.