Recently I tried to introduce the readers of Der Standard to the wonders of procedural generation with this article, but VGT-colleague Christof Zurschmitten plunged the further depths of this topic in last year's article for games publication GameStar. During his extensive research, he conducted a series of fascinating, in-depth interviews with a few notable game designers on procedural generation in games today. In this interview, the first of a series of diminuitively nicknamed "footnotes", Christof talked to Arcen Games' Chris Park. Warning: long read ahead.
Your first game, AI War, as well as both of the Valley Games had a strong emphasis on procedural generation, so much so that I figure it has become a part of the image of Arcen. My first question would thus be: How did you get to focus on procedural generation? Was it because you were a programmer interested in the challenge that came with it or was it more a means towards an end for you?
I think that in my particular case, I didn’t become a game developer. I was already a professional programmer and game development was a lifelong hobby of mine. But I never thought that was going to become a career and so I’m a little bit of an odd one, in terms of how I started out with things. All that said, my reasoning for using procedural generation wasn’t motivated by anything market-related or trying to accomplish anything for other people. It was all focused on what I wanted as a player for myself.
"I wanted to have a computer opponent which felt more human, which felt more interesting in that it had replay value"
The reason I started creating AI War in the first place was that strategy games in general are all about the opponent and how the opponent reacts on you. I dislike playing against other people in strategy games, because they tend to get into very much of a rush, where everybody is just trying to see how fast they can be. I don’t find that very interesting. And so I wanted to have a computer opponent which felt more human, which felt more interesting in that it had replay value. And so, even though as a developer I was creating it, I was creating it for myself, so that in spite of the fact that I know all the rules and how the games work internally, I wanted it to be able to surprise me. In order for it to do that, it had to have a large amount of procedural work. So I really came at it from an AI perspective and procedural galaxy creation and all that sort of thing was really a by-product. It wasn’t something I intended to do; it was just something that fit so well with what I was already doing with the AI.
A lot of strategy games used to have procedurally generated maps, like Age of Empires II and so forth, but then they really tuned that back with Age of Empires III and the other, more modern strategy games. That really bothered me, because I really thought that this hurt replay value. They said they did it in the service of greater artistic fidelity, or whatever, having models for the trains that look better and stuff. I really put a premium on replay value, and I feel like as a developer or as a player, the only way that you are going to get open-ended replay value – not just high replay value –, like in StarCraft for example, which can still be played ten years after the fact that it came out, is to have some sort of procedural generation, if you’re not having player vs. player.
There is a popular notion that procedural generation calls for a specific kind of developer, one that is maybe closer to an engineer or a programmer than an artist in the classical sense of the word. Do you consider yourself an artist?
I am a little bit of an odd one, because I’m very much of an engineer, and I’m also very much of an artist. When I was going off to university, I was really split on whether I was going to get into creative writing or whether I was going to do computer science. I finished up taking lots and lots of both sorts of courses, so I have really strong artistic interests, and really strong engineering interests. For me, the procedural generation is a very interesting engineering challenge, that is true. And there is almost some sort of artistry to that engineering, because it’s like a really elegant math equation, I suppose, like what theoretical physicists do. It’s not like a routine-thing. You’re making something that is almost alive in some ways, and it surprises you, like something that is alive. (It’s not alive, but it has kind of that feel to it. So there is a fun artistry to that kind of engineering, but at the same time, it takes a lot of engineering to actually pull that sort of thing off properly. For me, from an artistic standpoint, it lets me go back and forth between the engineering and broader artistic or design concerns. (I think of more of design than artistic considerations.)
But I don’t have to focus on every last piece of level creation. I can look at the forest instead of the trees, so to speak, with the design elements and let individual trees and shrubs get taking care of by procedural generation. But then I still have to look and see if a tree is upside-down, or if that shrub is bright pink – and they I say: “Oh, it’s time for some more engineering! Let’s fix the procedural generation.”
You said that you liked the fact that you are surprised as a player by procedural generation. Did you also have moments when you were surprised as a developer by something the algorithms came up with, something that you could use then for the development of the game?
Yes, definitely. In Valley 1, we used a variety of different cave digging-algorithms. If you imagine that you have a bunch of dirt, and it’s just kind of sitting there, broken up into smaller pieces. And you have got a little digger-guy who moves through there, digging caves. Every so often, he puts down a friend, and that friend also starts digging, and so forth. Using an algorithm like that, combined with some other rules, like dirt they can’t dig through, we round up with some really surprising shapes. So we said: “That’s totally unexpected. So if we give it this range of input, it gives us caves like these.” So we saved some of them, like the ones full of stalactites for the abandoned towns, or we used the caves that were really big and bulbous for the deserts, things like that. As we discovered different kinds of magic number inputs, we found that the same basic algorithms gave surprisingly different results, and we used those differences for different parts of the game, to give them a different feel.
Roger Ebert thought that the main reason why games could never be art was because they were interactive, manipulated by a player, and so contrary to the authorial intent he thought to be at the heart of a work of art. Framed like that, procedural generation is a funny thing: On the one hand, it’s an aspect that is really typical of the medium, something that opens up possibilities that no other medium has. On the other hand, procedural generation really basically laughs into the face of Ebert’s idea of “art”. Have you ever come across somebody with that kind of attitude, somebody who thought that what you are doing is not really of much value, because in the end, it’s the computer which does everything?
I am familiar with Roger Ebert’s stand on thing. I always liked him as a film critic, but his views on video games are pretty amusing, to say the least. I have never run into anybody with that sort of perspective personally. The thing that I think underlies a statement like Ebert’s is a lack of understanding of exactly what is going on to make a game, or what exactly is involved in something like procedural generation.
"There’s a very big difference between the word ‘procedural’ and the word ‘random’."
There’s a very big difference between the word ‘procedural’ and the word ‘random’. With ‘random’, you’ve got just anything. You have just noise. If you have, for instance, a random distribution of dots, you’re just going to get static. Eventually, everything will just fill up with dots, because every point has an equally big chance of getting a dot. That’s not procedural generation, though. Procedural generation is all about intent of the author. What that does is: It says, ok, we want this for this range of things to happen.
I don’t know if Roger Ebert is aware of this, but if you look at, let’s say, the movie Inception by Christopher Nolan – that has quite a bit of procedural generation in it, actually. There is a section where they have crumbling skyscrapers – there are a lot of them, and they all have to look different, they all have to behave realistically, the have to crumble realistically and so on. So what they had to do was use a program to do, procedurally, the 3D for that. Christopher Nolan is of course the director of this film, so he is like: “I want the buildings to look like this, I want them to crumble in general like that, and I want them to look realistic.” So the artists tell him how to do it in a way that is ultra-realistic, and so they use procedural generation. It’s not the fact that the artists on Inception were lazy, or that they wanted to have five artists instead of two hundreds animating those buildings.
I think that there is some point where you cannot just throw more man power or more people at a problem and actually get as good as a result, only faster. So they were able to achieve something with the collapse of those building that simply could not have been done any other way. Or, if you look at the battle scenes in the Lord of The Rings-movies, it’s the same sort of thing. They had procedurally generated AI for those and they were directing the overall flow of the battle, but what specific orcs did was all procedurally generated, so that it would look realistic. Of course, you can’t just hire 10’000 extras and put them all in costumes. So it’s not just a budgetary concern there. Things just stop working once you get a certain number of people involved, trying to pull something like that off.
I’m not a programmer myself, but I sometimes get the impression that procedural generation is a mirror image of game design itself: You define the rules, and then you observe what is possible to do inside those rules.
That’s a really good analogy, actually! It’s almost like the computer is the player and it takes whatever choices it can in the bounds that you define. At that point, you see what the player does, and adjust the rules accordingly, and observe what the player does, and so on.
So, let’s talk about A Valley With No Wind. In your reflections on version 1.0, you wrote that there were several points where you had the impression that the perception the player had was different to what was actually going on in the game. Do you think that this is also a by-product of procedural generation? Is it more difficult to subtly guide players’ perception in a game with heavy procedural generation?
It’s definitely possible to do something like it with procedural generation. The AI in AI Wars shows this. I read a lot about psychology, and I use that a lot in how I work with procedural generation, especially for AI. Because what people have to realize about movie making, game development, whatever, is that those are all basically magic acts. If you’re thinking about a movie, obviously outside the frame of the camera, there’s other stuff going on, people with microphones and that kind of stuff, and the actors are just actors. It’s the same thing with games. Behind those wireframe models, there is just emptiness. There is nothing there. If you see a hill and you can’t get passed it, it’s because there’s nothing on the other side of that hill, despite of the fact that your brain is kind of intuiting that there are more hills over there. With procedural generation, randomness is pretty much a helpful thing.
If you look at Valve’s Left 4 Dead 2, there’s a great example of it. When the zombies suddenly come up with a big surge, for instance, sometimes it seems like they are coming right when it’s the worst possible time for you, and you think: “Oh, that AI director was so smart!” But it’s probably not as smart as the players give it credit for. But the interesting thing about human brains is that we are built for building stories out of our environment. We try to make sense of what’s going on. We see some random events, we start looking for patterns, and then we build a story out of those patterns. It’s not a thing we consciously do, it just happens. So, in AI Wars for instance, when the AI randomly chooses some place to attack after the player has done something damaging to the AI, the players often attribute emotions to the AI. “That was a revenge attack!”, or “I made the AI mad!” They routinely attribute these human emotions to the AI. The interesting thing is that I do the same thing.
"We attribute human-like emotion and response to something that is really very much procedural and not human at all"
I know for a fact that there is no code for getting mad and taking revenge on the player. But when the AI seems to be acting mad and seems to be taking revenge on me through a random event, I think the AI is mad. And when I do something damaging to the AI and it does not do anything that constitutes a revenge-type behaviour, then my brain immediately goes: “I damaged him and he’s withdrawing into himself! I want to see what he does now!” So we attribute all this kind of thinking and human-like emotion and response to something that is really very much procedural and not human at all.
So, procedural generation can be an advantage. If people play through a scripted sequence twice, they immediately see that it is scripted, and they find it to be boring and unrealistic. You basically start seeing the puppet strings. But with procedural generation, it takes much longer to see the puppet strings, because of how different it is each time you play through it. Eventually, you start understanding that by doing something specific, the AI will react in, say, one of five ways. But it takes a while before you start figuring that out, you have to really be watching for the pattern to even figure that out.
What you said explains, for example, the popularity of all those Dwarf Fortress-diaries. It’s a system that is so complex that people still do not understand the patterns.
That’s a great example. Dwarf Fortress is such a cool game.
A lot of people are saying that the future for narrative in games is either something like DayZ, where the interaction of the players between themselves leads to emergent storytelling, or games like Dwarf Fortress, for the reasons you explained. Do you think the future for narrative in games should be going more into that direction instead of throwing more and more resources at scripted events, stories and so on?
I don’t think there is any one answer to that question. I think diversity is really good, and I think it is interesting that things like DayZ and Dwarf Fortress work. Those are really difficult to make, and they are new and exciting. I hope we’ll see more games that do things of that nature. But a) not everybody has the engineering to do that sort of thing, and b) I really would hate to think that any one method of storytelling will completely replace all others. I really think that diversity is one of the prime aspects of our medium. It’s like all that we are going to have are superhero movies from now on, that studios would be like: “You know what, we’re not going to have any more romantic comedies, because superhero movies are so much better!” That doesn’t make any sense. Of course, we’re going to have both. They serve different audiences and illicit different things.
Arcen is known for putting an exceptional amount of care and effort into post-release support. Do you think that it’s different to do that kind of support when you have a game with heavy procedural generation?
It depends. Both have their pros and cons, I don’t think that one method is overall easier. The next game that Arcen is doing – Shattered Haven – has no procedural content whatsoever. Shattered Haven is something we have been working on since 2008; it’s something that actually predates AI Wars. My feeling after having worked on both kinds of games is that you spend the same amount of time in either case, but you work on different things. With non-procedural games, you’re likely to have less replay value in some senses, but an overall tighter experience. That does not really apply to strategy games, though, because they are all about the foes going back and forth. But if you’re talking about adventuring, I think a certain amount of hand-crafting is necessary, which is why there is a lot more handcrafting in Valley 2 than there was in the first game, for instance. I think you wind up spending the same amount of time either way.
Ultimately, I don’t think that procedural generation saves you any time at all. But I think it allows you to create different experiences with the time you do spend on it. Experiences that do have a lot more replay value.
Since you chose to use more handcrafting for Valley 2, do you think that this will lead to a better experience, or just to a different one? Did you have the feeling that the first game was maybe a bit too chaotic, incoherent or random?
The first game really was a mix of a lot of things. It was a little bit eclectic and therefore something that appealed more to an edge than we had intended. Enemy placement is a good example – there are just a lot of things that you can do with handcrafting that you cannot do with procedural generation, just because the rules get so complicated, and there are so many rules. One of the things about procedural generation is that, the more specific you get with the rules, the less procedural it is, and therefore the less interesting. So you are going to have some kind of outlier funky stuff that happens if you relax the rules too much. But if you tighten them too much, then you get stuff that is really bland and uninteresting. So you have to strike that balance.
I think that we did that as much as we could with Valley 1, and that it came out as good as it could possibly have been. With Valley 2, what we decided to do was take what we call ‘slices’, which are blocks of handcrafted content, and then we are procedurally combining them and rearranging them in various little ways. We have about a little over 3000 slices that were created by our stuff and our contractors for Valley 2. That’s how we do the terrain generation in this game. It leads to millions of possible combinations. You wind up seeing some stuff that will make you think: “Oh, I recognize that terrain formation, I have seen it before.” But it’s still more interesting than just pure procedural generation. We have been learning as we go, and we have been treating it as a kind of research project on what is the optimum for what is the optimum of blending procedural and handcrafted content. I don’t feel that there’s one single answer to that, but I much prefer for myself the blend in Valley 2 compared to Valley 1.
When I was a child, I dreamed of the “infinite game”, something that did not really exist on consoles. What was your main inspiration for the Valley-games? Did you just set out to do a ‘procedurally generated Metroidvania-game’, or was there another spark that started the project?
We wanted to do a never-ending game – it’s that same dream of childhood, a game that is infinite. That’s something that Valley 1 attempts, and, I think, succeeds at as well as it can be done at the moment, at least with our budget. Valley 2 steps from that and makes a game with infinite replay value, but without each world being infinite. So the focus is different. But originally, the goal was to make an infinite world that you can go and explore, find stuff in or adventure through, and continuously power up. We felt like we did as much as we could with that concept with Valley 1, but in the meantime, we had made all those mechanics that we felt like we could do more with, in a more directed fashion. So we made Valley 2, where it’s more of a ‘you vs. the forces of evil’-thing, with a defined narrative arc. It uses procedural generation along the way to help giving you the infinite replay value and a lot of surprises as you go.
So let’s talk about the historical perspective a bit. Games with procedural generation have been around for a long time, Elite, Daggerfall and so on. But lately, we have been seeing a sort of renaissance. There are all the rogue-like crossover-genres, there are the endless runners on mobile platforms, and so on. Would you agree that there is a renaissance taking place, and if so, what are the reasons for it?
I definitely do think it’s growing in popularity. That said, you can go back much further than Daggerfall – for example, there was Adventure for the Atari and other games that had procedural content. It really has been around since videogames first existed. Back in the 80s, procedural generation was done largely because of the memory limitations. You did not have enough room to store all the levels they might create, so they had to build them out of procedural generation. The reason largely was hardware limitations back then.
"Procedural generation fell out of favour because people wanted to show that they could design with intent"
I’m speculating heavily here, but I think that there was a big shift in the late 80s and into the 90s, where procedural generation fell out of favour because people wanted to show that they could design with intent. They had ridiculous amounts of storage space, and wanted to show all the things they could make. Procedural generation was tossed aside out of those concerns, that attitude of: “Wow, look at what we can do now!” It’s the same reason that 2D stuff was tossed aside in the middle 90s. It was not that 2D was suddenly bad. It was just that everybody was so excited that they could do 3D that it was all they wanted to do for a while. It took Nintendo until 2006 before they did New Super Mario Bros., which was their first proper 2D Mario-game since the early 90s. So it’s not just Indies, and it’s not just a matter of budget that these things – 2D and procedural generation – come back into favour.
I think it’s a matter of developers getting comfortable with the fact that really hi-fi 3D, specifically scripted set piece type of games are possible, and they are cool for what they are. But they are not the end-all answer. That maybe sounds almost derogatory, but that’s not how I mean it. For a procedurally generated game, that’s true as well: They are cool for what they are. They are what they are, and nothing more. We are seeing a situation now where broadly speaking, game developers from AAA down, but in particular in the indie spectrum, are pulling from all of gaming history, and using all of the tools at their disposal now, rather than feeling like they have to use only something that is trendy. The definition of what is trendy has really expanded to be all encompassing of all of gaming history, and not just on what only the latest hardware can do. There was a time when a good game had to be a good game and a tech demo of your latest hardware, and that’s no longer the requirement. [laughs]
All that said, one still gets the impression that procedural generation is largely limited to certain genres. Do you think that there are other genres that could profit from using it more prominently, or genres that do not exist yet, but could exist by using procedural generation?
I think very much so. I don’t have any specific ideas on that at the moment, or I would probably already be working on them. But I like to find the space between existing genres anyway, and do things that are different. If you look at A Valley Without Wind, for instance – Metroidvania games in general are defined partly by their really tight design that includes backtracking and a lot of really tight level design. We made a procedurally generated Metroidvania game, which means inherently that it does not have that handcrafted tightness. So Valley is arguably not quite a Metroidvania game, it’s something else.
I think that there are a lot of genres you could do that kind of thing with and you will get really different and often interesting experiences. You don’t just get more of the same, game 47 in that franchise. You get something that is actually genuinely new and brings to life some new ideas. I think there are all sorts of games that could benefit from that. I don’t think that procedural generation is a magic bullet, but I do think that it’s a really interesting tool, and I doubt that we have found all the potential uses for it yet.
What role do you think procedural generation will play for the medium in the future, say, ten years from now? For example, I’m also interviewing Michael Cook, who is trying to come up with an AI that can design whole games for his PhD project. There definitely is a lot going on there. That said, do you think that procedural generation will become even more important in the future?
I think that there are types of games that do not need procedural generation at all. For example games that have a really tight narrative, that are all about the story. There you really want to control how the player feels and how they proceed through the game. If you are basically making an interactive movie or novel, in other words, you do not need procedural generation. It is likely to weaken the story. Something like that is really not going to see any inroads for procedural generation at all. Outside of that particular realm where the story is the end-all purpose of the thing that is being created, I think that you are going to see some degree of procedural generation in pretty much everything, because it helps with replay value so much. Right now, AAA-games are leaning on multiplayer for their replay value – every game must have multiplayer.
But not every game is suited to PvP-multiplayer. In the absence of that, procedural generation or intelligent AI is instead the answer for the replay value. I don’t think that every game ever is going to have it, but it’s a tool that is coming back into use, because we have set it aside for a long time. I think that we’re seeing maybe 30% of the capacity of peak usage of procedural generation. I think it’s going to become more and more popular, and eventually hit a point where we will see a backlash, where it will fade from popularity. Ultimately, though, it will stabilize somewhere north where we are now. That would be my guess, if I had to predict.
It’s interesting that you say that. So far, you get the impression that the AAA-sector uses it in asset creation, but they often try to keep it out of the sight of the player. It’s rarely used on the gameplay level.
That is my impression as well. It’s mostly used in a lot of pipeline asset-creation. I was really excited that Valve decided to have procedurally generated changes to their levels in Left 4 Dead 2. I hoped that there would really be roads that were open sometimes, sometimes not, and so on. Not to criticize them, I mean, they at least did it at all. But the way that they did it… there were only two or three points, some doors that might or might not be open, but if door A is not open, then door B around the corner will be open. So it does not really make much of a difference. I was a little disappointed in that. I thought that while what they did was cool, what they could have done is a lot cooler. But honestly, I would be really surprised if in Left 4 Dead 3, they will not have something that is a lot more morphological, so to speak. I think that is coming.
So, one last question: What does the future hold in store for Arcen Games? Valley 2 is not officially published yet, was it?
The release of Valley 2 is our next main thing. Then Shattered Haven is coming to beta late this February, and it will be released in March probably. That one is a lot more story-focussed, it’s all about handcrafted puzzles and solving environmental puzzles. So it completely steps away from procedurally generated content at all. There is none of it whatsoever.
Was that like a cure for you after all the procedural generation you did?
It’s more of a change. As I said, it actually is our oldest game, one that I started with before we ever did any procedural generation at all. It’s a game I always wanted to come back to, but for various reasons had not. I like making different kind of things. I don’t really want to become known for any one thing. So I felt that the timing was right after having so many procedural games to do one that is not.
The next game after that is going to be really story-focussed. It’s kind of a horror game, which is much darker than any game we have ever done before. No procedural generation whatsoever, and a really different kind of gameplay that is much more narrative-focussed, much more of a shorter game, with hi-fi, 2D art. It is a big departure for us which wanted to do to show that we can do different things. (But, of course, also because we had an idea for a project that we thought was really cool.) And then we have got a bunch of other stuff planned for along the line, some expansions to our existing games, some other strategy games and so forth. We also have planned several games that use procedural generation again. We are by no means giving up on procedural generation, it’s going to be probably in at least ¾ of the games that we do in the next years. But we don’t want to be limited to it exclusively.
Good luck for your future projects and thank you for your time!
Addendum: This interview was conducted in 2013. Since talking to Chris, the impressively productive studio has released a whole string of games: The already mentioned A Valley Without Wind 2 and Shattered Haven, but also the "turn-based 4x simulation god-game" [sic!] Skyward Collapse and Bionic Dues, which was smart enough to pair the roguelike genre with Mechs, and use, again, procedural generation to great effect.
Check back later for interviews with Jim Rossignol (Big Robot and Rockpapershotgun), Pwnee Studios (Cloudberry Kingdom), Jake Solomon (XCOM: Enemy Unkown) and Michael Cook, who is developing an Artificial Intelligence which makes games (and one day will probably rule us all).