So... Tynan; Emergence?

[Vagabond]

Tynan,

I understand you're a very busy man, so don't feel obligated to respond if you don't have time to sort through my dribble.

Basically, I want to probe your brain for information regarding emergent gameplay; Emergent AI to be specific.

A game from one of my favorite series is being developed at the moment by GolemLabs called The Guild 3. They've written some interesting articles on their form of Emergent AI, called "Evolutive Human Emulator", or EHE. Just for reference, here are the links to them:

Here,
Here,
and here.

I know you attribute your AI Storytellers to Left 4 Dead's AI Director. Apparently Everquest Next is working in colaboration with Storybricks, using their Emergent AI engine.

Is the emergent AI the future of AI. Will it come to a point where it is simply the new standard for "the computer"? Especially given the fact it adds so much to the feel of narrative, newtonian law, and emotion.

I know there is much more to it than is covered by those articles, but what I'm really interested in is exactly how different (on the programming level) is an emergent AI from one that isn't emergent, within the realm of game engines.

Cheers,
Michael

[Silvador]

Wut... You think the storytellers are some kind of advanced intelligence? More than likely they are nothing more than another level of the "difficulty setting" in which one can control at what pace their settlement is reduced to rubble via RNG.

It's really not hard to figure out. One gives a truly "real" feel by offering no control (Randy), another offers a chance to focus on building, as some players prefer to do, with minimal fear of having to defend one's construction on a repeated basis (Phoebe), while another allows for a more typical, gradual and steady increase in difficulty, giving one a chance to grow accustomed to their situation and establish a footing before being bombarded with ruthlessness (Cassandra). On top of these is an "intensity" setting, allowing one to further fine tune the difficulty and danger of their experience.

At what point do these simple, pre-programmed settings become anything even remotely close to resembling some form of an actual AI? They're sliders. A simple set of controls that allow tailoring. They're given the title of "AI Storyteller" to avoid them simply being blandly labeled as what they are, in-depth difficulty settings.

[huyderman]

AI stands for Artificial Intelligence, not Advanced Intelligence. And the storytellers are definitely a sort of simple AI. With the possible exception of Randy Random, they make decisions (e.g. which kind of events to fire) based on "observations"/inputs (e.g. things like the worth of the colony and how long you've been playing). And there are many similarity with L4Ds AI director which decides when to spawn zombies based on the state of the party to create lulls and peaks of intensity.

Sure, neither are Skynet, but it's definitely AI, as the term is used in games.

[Dr. Z]

I think cassandra is skynet 

[Silvador]

AI stands for Artificial Intelligence, not Advanced Intelligence. And the storytellers are definitely a sort of simple AI. With the possible exception of Randy Random, they make decisions (e.g. which kind of events to fire) based on "observations"/inputs (e.g. things like the worth of the colony and how long you've been playing). And there are many similarity with L4Ds AI director which decides when to spawn zombies based on the state of the party to create lulls and peaks of intensity.

Sure, neither are Skynet, but it's definitely AI, as the term is used in games.

They're no more "intelligent" than the raiders. They don't think, they don't "decide". Variables are programmed in and when certain criteria are met, the either or logic of the program spits out an event based on those criteria.

If A is true, then roll RNG to select from bank of F events.
If A is false, move on to B.
etc.
etc.
etc.
etc.
etc.
etc.

It's basic programming, built on numerous layers to add complexity. Our actions directly influence what gets thrown at us with a little RNG thrown in.

The term Artificial Intelligence, in the case of gaming, merely refers to a program with complex enough variables given to it in order to provide the impression of something other than predictable, basic patterns.

You'd probably find more intelligence in CoD's Fish AI.

[Vexare]

Did you guys read any of the articles the OP linked? Particularly the second one gives a better idea of the game-maker's approach to "AI" which is just a buzzword you're referring to in game technology and how a computer decides 'emotional' responses based on scripts and algorithms.

I don't think he's referring to the storyteller's AI which determines the frequency or predictable behavior of attacks and 'events' that occur in the game. I'm thinking this is more about the behavior of the colonists.

In RimWorld, colonists and pirates and tribals all have 'traits' which give them innate tendencies. Those are all scripted outcomes that when mixed together determine whether the 'person' (pawn/unit) will break sooner than others in a stressful situation or how fast they move, or their overall moods and motivations. This is the 'sim' portion of the game. Emergent AI deals with that a bit more in gaming where personality quirks come into play and are factored through a series of possible outcomes that could occur. Such as, in the case of a mental break ... what plays into determining if the colonist just wanders around aimlessly for awhile and then snaps out of it, or goes berserk and tries to kill everyone? Does their 'personality' and traits decide that or is it simply a random roll of the storyteller's AI that decides?

Vagabond is wondering if Tynan is tapping into deeper AI ideas or leaving it to a simpler random roll type functionality in determining outcomes for situations in RimWorld not just for 'events' randomly (or predictably) generated by the storytellers but also for the individual colonists actions and responses.

[Vagabond]

Vagabond is wondering if Tynan is tapping into deeper AI ideas or leaving it to a simpler random roll type functionality in determining outcomes for situations in RimWorld not just for 'events' randomly (or predictably) generated by the storytellers but also for the individual colonists actions and responses.

Exactly this.

The difference between even Rimworld's colonists and NPCs in just about any rpg is vast simply because there is an emotional state based off of world variables. Then you have Steve Grand's Creatures/Grandroids and Rimworld's colonist; another big leap between them with Grand's creatures being things with evolving ability to understand as they age, ability to learn new things, and ability to start or stop doing things they have/haven't learned.

I suppose the "proper" term would be VI rather than AI. Considering the way they are defined, with AI being self aware, I don't want anyone to think I'm under the impression that any of these games are somehow having a magical digital human in their processor. I'm not delusional, I'm just facinated with the intricacies of virtual intelligences. I understand that it is all programmed behaviours, but the method in which the computer determines what happens next is vastly different than games which don't elect to program their game in this way.

Once you begin to code emotional states, personalities, and genetics, they cease being simple moving stat blocks with specific lists of actions to be done at certain times based on priority of task, and become emergent virtual intellegences.

Take a colonist who has firefighting disabled. Wouldn't it be so much cooler if it wasn't simply disabled, but something that simply scared them; they would see the fire and become unable to function normally, given the urgency of the situation of a buildign being on fire, they could attempt to put it out given a normally high level of bravery, but their resolve would flounder as they went closer to the fire, till they began to cry or feint from anxiety. What if they were somewhere where fires occured often, is there a way for the colonist to get over this fear?

It is far more evolved than say an npc in Skyrim. They are just... there. They have a specific list of conversation topics, a specific list of places to path to and when, and no real personality or emotion. This is the standard "AI" I was talking about, versus the more complex one.

I don't know... I just wanted to see what an active developer thought about it.

Cheers,
Michael

[Silvador]

"tapping into deeper AI ideas"

I... I'm just going to go, now... and leave you... people... to.... whatever it is you... are thinking. Goodluck with that.

[Tericc]

@Dave

As long as we don't get a story teller named Hal, I think we will be ok?

[Tynan]

I've specifically not used any indirect or "emergent" styles of AI programming. They're too hard to predict, too hard to debug, too CPU-inefficient, and too inconsistent to be useful.

These sorts of projects have been talked about incessantly for 40 years, yet they're still not really useful for anything. To me that's an indication that something is wrong with the idea at a fundamental level. When I see devs trying to apply strong learning AIs or neural nets or some sort of "generalized" adaptive intelligence engine, I just cringe. This sort of thing still belongs in academic research labs, not real games.

As for the word "emergence" itself, it's kind of being abused here. Emergence just means complex large-scale behavior from simple small scale pieces. A puddle of water has emergence. And so does the mechanistic AI of games like DF or RimWorld, when you put together a bunch of actors and things in a scenario and let a player watch and interpret stories from it. Emergence has always been a cornerstone of the RW design.

[Kegereneku]

I will resist the urge to get on another debate about Emergent Gameplay with Vagabond. I will resist the urge to get on another debate about Emergent Gameplay with Vagabond. I will resist the urge to get on another debate about Emergent Gameplay with Vagabond. I will resist the urge to get on another debate about Emergent Gameplay with Vagabond. I will resist the urge to get on another debate about Emergent Gameplay with Vagabond. I will resist the urge to get on another debate about Emergent Gameplay with Vagabond.hope Tynan and other will correct his terminology and abuse of the world Emergentthat isn't equal to AI as well as the need to knew the stuff well for it to give a good result.

Though, I'd love to ask Tynan how he'd use the term 'Emergent Gameplay' as the definition I was most familiar with warped toward something global or hype-generator.

** head explode **

[joeloverbeck]

Oddly enough I was thinking this morning about posting a similar thread. When Tynan talked about changing the behavior of raiders, so they would avoid killboxes and in general learn to siege and assault a fortress better, it seemed to me a good problem to evolve offline with a genetic algorithm. Set up a fortress, some regular colonist AI well armed. As a fitness function simulate populations of chromosomes that represent different orderings and configurations of a behavior tree (if you are actually using one) and the weights of utility functions (again, assumptions). High scores for those configurations that damage side walls, inner buildings, kill most people, stay alive longer, etc. After some (likely plenty) generations, could end up evolving very interesting stuff.

A classic example of that was in the first AI Game Programming Wisdom book (the "Evolving the Perfect Troll" article). I'm mostly interested in those kinds of crossovers of academia and fun, so I don't know how useful they would be for an accomplished game programmer. However, those kinds of tunings of offline parameters and AI configurations would seem very likely to benefit from that category of machine learning methods.

On a personal note, I was waiting for some great Dwarf Fortress-like game to appear and this is the best by far. The other concept I regret the industry didn't evolve was that old mid nineties "Creatures" game.

[Vagabond]

Oddly enough I was thinking this morning about posting a similar thread. When Tynan talked about changing the behavior of raiders, so they would avoid killboxes and in general learn to siege and assault a fortress better, it seemed to me a good problem to evolve offline with a genetic algorithm. Set up a fortress, some regular colonist AI well armed. As a fitness function simulate populations of chromosomes that represent different orderings and configurations of a behavior tree (if you are actually using one) and the weights of utility functions (again, assumptions). High scores for those configurations that damage side walls, inner buildings, kill most people, stay alive longer, etc. After some (likely plenty) generations, could end up evolving very interesting stuff.

A classic example of that was in the first AI Game Programming Wisdom book (the "Evolving the Perfect Troll" article). I'm mostly interested in those kinds of crossovers of academia and fun, so I don't know how useful they would be for an accomplished game programmer. However, those kinds of tunings of offline parameters and AI configurations would seem very likely to benefit from that category of machine learning methods.

On a personal note, I was waiting for some great Dwarf Fortress-like game to appear and this is the best by far. The other concept I regret the industry didn't evolve was that old mid nineties "Creatures" game.

I agree. As a student working on my bachelor's, I kind of have this notion of "I wan't to be the one to do it". Realistically, I know I'm no where near the ability of industry veterans. Failure is inevitable, but the ride could be worth it. If everyone just throws up their hands and goes "It's never gonna work"... Well.

The addition of mental states, hunger, and weather... This game is already emergent. Tynan defines it much the same way as Peter Molyneux, technopedia, Warren Spector, Richard Garriot, and others. It is simply adding more and more layers of complexity , wherein the character stops being pre scripted to do actions, and instead does them for reasons involving their state of being (Not that they are some advanced form of AI-Skynet..  ).

Take a look at... Oh.. Skyrim. Good rpg, but in comparison with Rimworld, all those characters are so lifeless. They don't have routines (that aren't prescripted, go here at this time, unless there is this weather condition, in which case go here and do this). They don't have needs, nor desires, nor a model for health. Imagine if they did, though?

The emergent narrative, or storyteller/director is a product of these variables. When the game determines what to do based on the state of the actors; the players progress and the current state of psychosocial behaviors occuring at various intervals of time... Well... This is emergence. The intentional emergence that comes from complexity, not from glitches in programming; not the emergence that comes from exploit or flawed design.

One of the things I'm eager to see is an expansion of the backgrounds and traits to include more in depth personality modifiers. I touched on this somewhat in my "Expanded Traits" suggestion thread. In this, relationships could evolve between colonists, with (I think) amazing results. The storyteller could use these relationships to good use, such as by inflicting a colonist with a heart attack the moment a conversation pops two colonist from best friend to a couple. She has a mental break, and he's on the ground floundering. Save him in time and alls well. If not, he dies, and she loses someone very dear to her.

The storyteller could send someone to join your colony that will obviously not mesh well. Or maybe someone who will? Maybe the guy then send you is a complete and total prick, but if traits can be changed over time, depending on outside influence, that guy could eventually be an integral part of the colony.

As to the personal note: The Creatures series is one of my favorites. Steve Grand is currently developing "Grandroids" which seems to be the evolution of Creatures you might might be interested in- I've been following; Check these out:

Kickstarter
Steve Grand's Blog

Cheers,
Michael

[joeloverbeck]

Issues with using more advanced AI in games is that they can evolve to a point that the game isn't fun anymore. If, for example, a neural network that controls the enemies keeps training, apart from the processor time it would take from other functions even with proper threading, the result could be so challenging or optimized that it would make the game unplayable. In general too hard to predict. A game is also a balancing of everything it has to do and keep going at 60fps. AI takes a lot. Offline training/evolving methods can take from minutes to hours, which obviously doesn't make them very suitable for using them in a game loop. Still I think that they would be exceptional at building a proper game experience: evolving the neural network of an enemy so the utilities it generates for every combination of inputs result in reasonable actions for an expected game experience. For example, a maze monster that defends its children and treasure while fending off adventurers, in one case, or another monster that depends on the average fitness of several other monsters, so it evolves a collaborative behavior. As a programmer, you would do this beforehand, using a barebones simulation of the game as the fitness function. Just this afternoon I read an article about this on the Game AI Pro book.

I doubt Skyrim could do much more with what it needed to do. Huge world and great graphics doesn't leave much room for advanced AI unless you want the game to run at 5fps. Still, I hope some game developers come up with a population/world manager that, even if turn based (advancing an hour for turn, for example), focused on the most advanced AI simulation possible, and going into small roguelike, one character play when needed, using data from the global simulation.

[Sharp]

I like F.E.A.R's AI which uses GOAP (which was made for it) by Jeff Orkin

It could possibily be utilised for Rimworld for Raider/Tribe AI so they can try and intelligently create a plan while weighing up risks, given reasonable goals like steal goods or burn crops, maybe even advanced goals like "Capture X due to their skill Y"

I would like RimWorld factions to evolve to a stage where they have a limited number of resources to utilise and will then create objectives based on their needs, and also if hostile factions perhaps have a need of something and will fight to get it or even just demand it in exchange for not raiding you, but I digress.

Grandroids looks interesting, will definitely check it out.