Forum Announcement, Click Here to Read More From EA_Cade.
Q&A with Alleged Lead AI Programmer for Sims 4
simmerorigin
Posts: 1,401 Member
Full comment by user @ rezination replying to this YouTube video "The Genius AI Behind The Sims"
I was an AI programmer on The Sims Medieval (which was based on The Sims 3) and the lead AI programmer on The Sims 4. This is pretty spot-on! There's more to it of course, but the core of the AI system is just as described here.
Things like visiting other sims were called Situations which helped define some of those rules. They weren't hard-coded, we managed it all with buffs, mostly created by designers. Buffs can be thought of as tags and you could have tests on interactions that said "don't allow someone with this buff to do this thing". So if you visited another sim, you got a visiting buff which prevented you from doing things that were inappropriate. You could also be invited to stay over, which was a different buff that relaxed many of those rules. Beds were a special case; they were owned by different sims so you would never autonomously use someone else's bed unless you were very close (like romantically involved).
Socials on The Sims 4 were governed by a special autonomy mode called subaction autonomy, which was a much simpler weighted random compared to the complex utility system of full autonomy. Designers would tune each social and weights for things like reactions and which social was chosen. Weights were adjusted by traits, buffs, and the short-term context (the short-term companion to the long-term relationship score). Short-term context models things like "I love my wife, but I'm [angry] at her right now", though in practice it was more like "this conversation is tanking" or "this is a really funny conversation". For example, when you're laughing and joking around with your friends, you're more likely to laugh at jokes than if you're just told a joke out of nowhere. This is even more true if everyone is laughing (that's why sitcoms have a laugh track).
For things like venues (restaurants and so on), we used the same core system. It's a Situation under the covers. Meta autonomy pulls in sims to fulfill different roles based on filters. It tries to use existing sims but if it can't, it creates a new sim. Townies are sims that don't live in any specific lot in the world. Any sim you can interact with are fully simulated. The auto-satisfy curves are used when you, the player, enter a lot with a bunch of sims that are already there. We use that to determine starting motives. The low LOD simulation is very, very light. It's mostly just story progression (getting a new job, promotion, etc.)
One thing the video talks about at the end is the ambiguity of the AI. This was very intentionally designed. Sims need to be reactive and always living in the moment. So while you might have a sim with an ambition to be a movie star, very little actually causes them to do that. They might enjoy movies more, for example, but not much else. Sims don't plan in any way. They run an AI tick and choose something to do based on motives, traits, mood, and so on. We've experimented with planning, but it caused confusion because players didn't know why their sim was doing what they were doing. Playtesters felt like they didn't have as much control.
Simlish was created by Stephen Kearin and Gerri Lawlor, two incredibly talented improvisers. I recommend looking them up, they're both awesome! (I've been doing improv for many, many years and I've had the pleasure of working with Stephen on several occasions.)
Sims being gay or bi vs straight is not entirely for storytelling, it's because of other countries. The design is that if you never ever initiate any gay content, you will never see gay content. This helped us get around laws in countries like Russia. That said, the video is correct that your player-controlled sim will never romance another sim, though non-player controlled sims can get married through story progression. Still, you won't see two men or two women get married unless you directly make it happen because of this issue.
There is a sort-of yes, and idea in the sims, but it's less designed than you might think. Basically, we have a set of interactions that can never be done autonomously, like quitting your job, getting married, and so on. We have the player make the big life decisions while the AI takes care of the rest.
If anyone has any questions, I'm happy to answer them.
He/him | Simmer since Sims 1 | Active Sims 2 wants-based rotational player, Sims 3 legacy player | My gameplay rules via PleasantSims | Bring back challenge and depth to the Sims: https://forums.thesims.com/en_US/discussion/comment/17959464/#Comment_17959464
Post edited by simmerorigin on
0
Comments
I've been modding the game for years, and it's funny to recognize a lot of these terms from seeing them all over the code. It's actually a very neat example of how object-based programming can give you a lot of variety.
I do have one question, if you can answer it: By now (through leaks and datamining) we know the game was meant to have a multiplayer element on launch. How late in development was that scratched and did it affect the development of other features at all?
Sorry, I can't talk about that.
" The design is that if you never ever initiate any gay content, you will never see gay content."
Reminds me of those people complaining about seeing gay ads, only to be told "you only get gay ads because you search for gay things".
We once got a bug on The Sims Medieval where someone was freaking out because two male monks were cuddling and he swore he never did anything gay in the game. Me and the lead gameplay engineer looked at that and rolled our eyes. Sure ya didn't. (Spoiler: he did.)
I have a million questions. I got into game dev because of AI. But now that someone with legit credentials is here, I can't think of anything! I'll just sub you and maybe ask sometime in the future. Yay learning!
Feel free to ask me anything. Some questions I can't answer because of NDA, but anything about general AI I certainly can. It's been my career.
Why did EA never fixed the performance issues of Sims 3?
I'm not sure how to answer this. Because programming is hard?
The Sims is incredibly complex and there's a balance between feature requests, bug fixes, and performance optimization. The producers and execs determine where to spend time. It's also very fluid; certain lots are WAY more expensive than others, and the player can do whatever they want. We generally optimize for typical lots. There are also different bottlenecks. Are you CPU-bound? GPU-bound? Memory-bound? And finally, certain architectural decisions preclude certain optimizations. At some point, you'd have to rewrite 500,000 lines of code to get a few frames back.
I very distinctly remember that venue AI in the Sims 3! It seems like whenever you brought a Sim out on the town, it would cause every other Sim in the city to converge on their location in a swarm (even at 3am doing things they were bad at!). Understandable from a standpoint of trying to spark interactions but it was a little eerie sometimes.
Thanks for the peek behind the curtain!
That's pretty much exactly what happened.
I'm curious about the reasoning behind the neighborhood stories AI. What weights are taken into account to determine the likelihood of events? Do things like age, location, previous friendships, jobs, the recently added compatibility scores etc. influence what people get married, for example?
Oh man, it's been like 10 years, I don't remember the specifics and I didn't write that specific code on The Sims 4 (I mostly did the motives, full autonomy, and subaction autonomy). I'm pretty sure we did weigh things like relationship, so two NPCs with a high relationship were more likely to marry. The decision making was pretty simple because we didn't want to have a performance spike.
With the thousands of hard coded rules Mark mentions in the video, how are these balanced over time? I imagine adding rule number 3517, which felt very important in a certain context, must be hard to weigh against all others?
So the rules weren't if statements in code most of the time, they were tuning values once could change. But yes, balancing all of those tuning values is incredibly complex. On the Sims 4, our lead AI designer went dark for like a week just tuning fun-based interactions for the base game.
We had a QA team that knew the game really well helping us, we had automated testing and smoke testing as well. We had unit tests to make sure we didn't directly break anything. On The Sims 4, I started up a weekly autonomy meeting where we would literally just play the game. There were a couple of engineers, the lead AI designer, the AI producer, and QA all in the room as we played. Just playing the game is so critical. I also had a secondary machine at my desk that was running the game all day every day.
Why are Sims 4 Sims constantly washing their dishes in the bathroom sink when there's a kitchen sink right there?
Probably because the bathroom sink is closer, or maybe the kitchen sink is dirty, blocked, or in use. If I recall, it was mostly based on distance.
Wow that's so cool! I practically grew playing the Sims, since the first one. I was kind of a control freak at the beginning, so I would get crazy whenever they tried to do mad stuff (or just slack as a normal human being), but I learned to get used to it with time, as I learned to let go and accept improvisation in every other aspect of my life as well. I would often think of the Sims as a metaphor for that.
I loved Medieval and also felt a huge leap on TS3 with the traits thing and the open world with the places that seemed full of life and so autonomous. I also felt that on TS4 many behaviours were addressed, making the Sims seem more sensible to the environment and context, avoiding situations that were awkward in other editions, and also giving the player more control in some aspects. Does that make any sense? I'm also curious as to how the background system of other Sims in town worked, and about architecture of the object-oriented aspect. As a developer and game dev, I'm really curious about so many things!
Also, can you tell us about situations in which designers were surprised by how the players were using features in a non-intended way?
Awesome!
We spent a lot of time trying to get sims more aware of their environment. There were many ways this was accomplished, mostly notably with the posture graph. Imagine every posture in the game (stand at desk, sit on couch, stand at nothing, etc) with the edges representing the different transitions. Posture transitions can also be a pathfind to a location. Now say that a sim has a plate of food and wants to sit down to eat somewhere. We run a dijkstra search across that graph, with each edge weighted by various things. One of the weights is directly related to relationship with the person in any adjacent location, so you will generally prefer to sit next to (or near) someone you like.
Another big difference is that we had truly multi-state interactions. The Sims 3 & Medieval had some of this, but for The Sims 4 the interaction system was truly a multi-state HFSM, so you could be in multiple interactions a the same time. There were a lot of issues with not trying to be too greedy with AI, but it worked out extremely well.
From an OOP point of view, we had a component architecture with a loose inheritance model. There was an AutonomyComponent that had all the run-time AI state for a Sim, but the actual AI itself was run by a class called the AutonomyService. Internally there was a queue of AutonomyRequest objects, so a Sim would run AI by instantiating an AutonomyRequest and submitting it to the AutonomyService. The AI itself was time-sliced for performance reasons.
The AutonomyRequest had all possible parameters necessary to run AI. In the video, he talks about being able to include or exclude different motives. This is where that happened, so you could easily prevent a sim from doing certain things. These were typically governed by buffs. Everything used this, so the grim reaper was just running heavily constrained autonomy.
We were surprised constantly on The Sims, which was a joy.
Also Medieval featured a story mode which was so beautiful and refreshing. Can you tell more about how that worked?
Do you mean the quest system?
My Sims 4 neighborhood is a sharply conservative place, based on the social values of my 1980s suburban childhood. When people complained that Sims 4 is "too woke", I never found that my own neighborhoods behaved that way—sure, had to edit a few gender-ambiguous premade NPCs in Create-A-Sim to bring them into line with the neighborhood's overall sharply conservative vibe (indeed, my username on the Internet has its roots in the way I play simulation games—simulated people in games in general are religious and I am their god)—but for the most part the AI wouldn't generate an interaction that violated the norms that my player control established.
Nice to see that was intentional.
One of the core values on The Sims is that we want players to play their way and tell their stories. If you want a conservative neighborhood, we don't want to stop you.
My biggest question is related to washing dishes. When there are two sinks on a lot, one in the kitchen and one in the bathroom, how come so many people find their Sims frequently deciding that the one in the bathroom is the better sink for washing the dishes when you tell them to wash all dishes on lot by interacting with a dish? Something I vaguely recall from Sims 2, definitely recall from Sims 3, and frequently read about people experiencing in Sims 4. Are there any changes to the AI systems that could be implemented without redoing the entire thing from scratch that would help prevent that, or at least make it less likely?
I answered this above, but the reason is because it's mostly based on distance. The video talks about how things are advertised from objects. The flip side of this is that there's less awareness for things like what room it's in.
How did you find this video so quickly?
I teach part-time and one of my students sent it to me.
What benefit do you get by letting the objects "advertise" their benefits vs just having the AI decide what need it wants to satisfy and then search for an appropriate object? I've seen several videos speaking highly of this solution, but the AI seems pretty straight forward to code either way.
It actually does both. Here's a really high-level overview:
1) Decide what motive(s) to solve.
2) For each object in the world, gather a list of the ones that satisfy the motive(s) I want.
3) Score every interaction on every object that solves for the motive(s).
4) Take the top scoring interactions and select one with a weighted-random.
5) Push that interaction.
So there's one AI algorithm running. I've read a lot of articles saying that the AI is in the objects and it's not. The objects have AI data on them that the algorithm consumes.
Are we living in a simulation? After this video everything looks so plausible to me
Yes.
Was watching the video, honestly half expecting to see an interview with Rez and lo and behold you're in the comments! Thank you for all the fantastic gdc talks and game ai articles over the years! Between you and Dave Mark, I cannot get enough utility ai.
Thanks!!
Proximity is probably weighted a bit higher than the type of sink in that case. Or in the version of Sims you played there was no distinction at all between sink types.
This is correct, there's no variation between sink types. It's mostly just distance.
I just want to say that Sims Medieval is a really good game, to me its probably my third favourite sims franchise game after sims 2 and 3. I wish that it had more dlc but mods it is I guess
Me too! It didn't do as well, so it was canceled after Pirates & Nobles. One of my favorite things was the faction feud AI stuff I did on Pirates & Nobles, and the Romeo & Juliet easter egg.
the second comedian who created simlish who you call Jerri Lawler is actually Gerri Lawlor, who died in 2019 of non-revealed causes.
You are correct, I mispelled her name! Let me see if I can edit the comment and correct the spelling. I never had the pleasure of meeting her, but many of my colleagues were very good friends with her. She was apart of BATS improv here in San Francisco (which I'm also apart of). Stephen was just on The Mandelorean as a senator not too long ago.
Why does the AI in Sims 4 perform so poorly compared to AI 3?
Perform in what way?
So why does The Sims 4 feel so soulless and lifeless compared to the first three? There's practically no risk of serious failure states like there used to be, there's no thrill to it, it's more of a dollhouse than ever with not much "game" to compliment it.
In what way? You mean why don't Sims fail more often?
How do you guys manage feature creep in this code base ? In the video he mentioned how much work the introduction of moodlets produced up to thousands of units of work.
We had really, really good producers. We also try to introduce features that can scale significantly in either direction. For example, moodlets didn't actually take thousands and thousands of hours of work; the core system was fairly complex but then it was just all tuning. On the Sims 4, we called them buffs and they had very defined ways to changing behavior. You could multiply motive scores, add new motives, deny certain motives, and so on. This gets you a lot, especially when combined with the work of our incredible animators.
And finally, we have to cut things. The Sims 4 didn't ship with pools or toddlers, which was a HUGE decision. But we worked on it in the background and shipped them free when they were ready.
thanks for the answer. I can see some abstractions helping on the extension of code and testing, keeping the entropy low on the codebase, but still, that's a wild feature creep risk, interesting you guys managed it, and kudos to the senior team on making a good core for the other to rely on.
Yeah, it definitely wasn't easy! Feature creep is the killer of shipping a game, especially one like this.
I can't really answer that. I can say that The Sims 4 shipped with a few major missing pieces (pools & toddlers) but those were added later. This is much more of a design/production question than a programming question. Besides pools/toddlers, it never felt like a lack of things to do, except of course the feeling that always happens when switching from a Sims game with tons of expansions to a base game with none.
I got a question!! My fiance actually brought this up recently… why will sims finish eating at the dining room table and get up and walk upstairs to the furthest bathroom to wash the dishes? Its very weird…
It's based on distance, but not exactly perfect pathing distance. We can't run a pathfind to every single possible washing location, so we estimate it. If I recall, at ship it basically used crows-fly distance (so length of sink position - sim position), ignoring walls, etc. For sinks on another floor, it does a crow's fly distance to the stairs, then from the stairs to the target. This might have changed since shipping (I left a few months after we shipped base game).
"We've experimented with planning, but it caused confusion because players didn't know why their sim was doing what they were doing."
It might be interesting to have a more simulation, hands-off mode where sims do plan and act.
But it would indeed need some feedback, players want to understand what happens. This was a major gripe of Sim Earth, sim ant, & Sim City (all).
Such a system could bring a welcome sense of authenticity in MMORPG NPCs.
I completely agree! I'm playing around with this on my current (non-Sims and non-EA) project. The hard part is balancing player autonomy with agent autonomy. In other words, the player needs to feel in control of their game, but the agents (AI characters) need to feel life-like.
I find the same sex aspect very interesting! I dunno if other players hear this but my parents once told me that I became gay because I play The Sims... If the same sex aspect said in this comment is truly the case, The Sims doesn't introduce same sex ideas into the player but only mirrors what the players input in the save file!
Is this also the case in the past and other versions of the game, like in 3 or 2?
Yes, I'm pretty sure this is how it worked on Sims 1 and 2 (though I didn't work on those, so I'm only extrapolating). Sims 3 definitely worked like this since we used the same core AI code on the Sims Medieval.
Also, for the record, a game can't make you gay, but it CAN help you explore things that may or may not be accepted. We've gotten letters from people talking about how The Sims gave them the courage to come out as gay, which is pretty amazing.
you already said a thing or two to "AntonLejon" about the rules that are in place to balance everything and that the only way to see if they are in a good spot is basically to play them out over many hours.
is there anything you where implementing you can recall, that was so impactful that it threw off the balance very unexpected or the other way around you did not noticed at all as you where hoping to see which in turn took way longer to tweak than others?
or do all those values had a hard caps to prevent some mass events?
and where there any that you had to scrap entirely because they messed others up to much?
There were certainly hard caps, but I created those caps, so they got moved from time to time.
One of the big issues that happened on The Sims 4 was sims being "greedy" with interactions. Because we allowed sims to be in multiple interactions, but best possible case was to have them in as many as possible, but that's not realistic (hence why you shouldn't text & drive...) I created the concept of focus cost, which was a limiter to interaction scores when performing a particular action. So certain interactions could be tagged with a focus cost that would artificially penalize choosing other interactions, preventing sims from wanting to do too much.
I think one of the biggest wins for me personally was getting design to trust the AI. On the Sims 3 and The Sims Medieval, AI was only run when the Sims's interaction queue was empty. The Sims 4 ran AI on a timer (mostly) because we wanted to be able to continue adding more interactions when necessary. So you might sit down to eat and then while eating, decide to talk to your spouse. The problem is that if a sim's motives were mostly solved, they would bounce between interactions pretty quickly. A social interaction would score at 5 and a fun interaction would score at 5. We might choose social, run one single social action, then leave.
Interactions had the concept of "always guaranteed", which means the interaction would run until something specifically killed it. Design solved the problem by using this, so they would force the Sim to stay in a social for 15 - 30 sim minutes (or whatever). The bad news is that this short-circuits the AI, so then sims start motive failing and I get bugs that sims are motive failing.
What I did is told design to never mark an interaction as always guaranteed unless it REALLY needed to happen, like a sim dying. The real bug was that Sims were bouncing around too much.
This was solved in a few ways. First, when scoring another interaction, we subtract the score of any current interactions that would be cancel. Second, we have an opportunity cost multiplier on all interactions. This is what we multiply that score by when subtracting it.
So let's go back to our original example: social is scoring at 5 and fun is scoring at 5. We run a social interaction, gain some social, and now social scores at 3 and fun scores at 5, so the net score is 2 (5 for fun minus 3 for losing the social). BUT, social has an opportunity cost of 2, so now we have a fun score of 5 - (3 x 2) = -1. Anything less than or equal to 0 is ignored, so the sim stays in the social.
That's a long winded answer, but hopefully that's what you were looking for.
1. What's your thoughts on MCCC modifications to population rules, autonomy rules, culling, performance and so on? Do you think they work well in tandem with vanilla code? Basically any opinion on MCCC from coding perspective would be super interesting.
2. Just a minor question, but to what extent is the AI affected by routing and how does that affect performance?
3. It's well known that in Sims 4 sims can get absurdly obsessed with very particular activities - like playing BlikBlok on that BlikBlok machine, drinking water from a sink, doing pushups in random places, just to name a few. There's some interactions that trigger way more often than it would seem reasonable. (Someone even made a "Less Obsession" mod addressing a long list of them). Do you think this is some oversights or bugs linked to how these items are coded, or is this by design?
1) I have no idea, I haven't looked at the mods at all. I will say that I'm absolutely thrilled with the modding community in general. It takes this from being an EA game to really being the Sims community's game. The modding community also does a lot without having the same tools we had. We don't ship any of the editors and I'm pretty sure all the comments and things are stripped out (which is a shame, I'm a VERY heavy commenter so it would make autonomy way easier to understand).
2) The AI makes a decision to do X. Part of this is a distance cost, but we don't want to incur the cost of making a bunch of routing calls for everything we're considering, so we do an estimation. Last time I checked, it was crow's fly distance if you're on the same level. If you're on a different level, it's crow's fly to the stairs, then from there to the target. It might have changed since I left though. Once a decision is made, we run a posture transition, which includes an actual call to the routing system. The posture state was stored as a graph and we ran a dijkstra search to determine the best posture transition, which could affect where a sim goes (such where to sit to eat, for example).
3) It's a subjective call on the part of whatever designer tuned those objects. It's incredibly difficult to make it feel good in all circumstances. For example, if you have a single high-quality fun item and a bunch of low-quality fun items, they will definitely obsess over the high-quality one (wouldn't you?) For things like push ups or the BlikBlok object, it's a tuning issue. for what it's worth, I do remember a bug coming through before shipping specifically about BlikBlok object being way to fun....
At the end of the day, it's a combination of bug triage, subjective feel, and a judgement call. Also remember that you players play the game very differently than many of us. You get a much better gestalt of everything and while we do play the game just to play, most of it is trying to look at a specific issue.
Hopefully that answered your questions.
I have a question. How did you manage to make this system scalable. Adding more objects to the lot would make the system to iterate on more posible candidates. You said you checked the objects by traits. Did you have a map of the objects in the lot by traits or smt similar?
There are a few ways. All objects knew what motives they could solve, so we only had to go to the object level to determine if we could do something. We also had some inverted interactions where we'd choose the action first, then we'd choose the target. Viewing paintings or insane sims talking to random objects worked this way. It was a "self" interaction (i.e. the original target was yourself) which would then override the target when running the interaction with a random one.
AI requests are also queued, so only one runs at a time. Furthermore they were time-sliced, so it would run for some amount of time and pause while the rest of the frame happened, then pick up where it left off. In full debug mode, it could take a second for AI to finish on a big lot, though in release it rarely came up.
But yes, scale is a huge issue. We ran into a bunch of scaling issues and spent a lot of the last year working on those kinds of issues.
As a sidenote, Medieval really was the best Sims game ever. If only it got some more family systems and more expansions with furniture and such, it would be so absolutely glorious. I still mourn the game it never became.
Me too.
I understand if you don't know considering if I read this correctly you didn't directly work on sims 3, but I heard recently that if you direct your own sims to be gay, it makes it more likely that gay sims will "naturally occur" within story progression
My recollection from The Sims Medieval (which was based on The Sims 3) is that once you have a gay interaction, it opens the rainbow floodgates and story progression allows gay sims to occur in the wild. This was 12 years ago so I might be misremembering, but I'm about 90% sure that's correct.
A lot of questions about the sink problem, the problem is probably distance, littlemssam has made a great mod (ImprovedAutoChooseObjects) which pretty much solves the sink issue where the mod lowers the allowed distance so sims prioritize the closest sink.
I think the AI in the sims4 is pretty good overall, it feels like they prioritize some things wrong sometimes but I think that's because some objects or actions are having more weight in the game so sims are drawn to them. I think someone here on the forums claimed that sims can be better at taking care of their needs by being "trained", I would have liked to ask them if that's the case because that could have explained why some sims seems much better at taking care of their needs than others.
Interesting read, thanks for sharing.