Since I returned from AIIDE I’ve been thinking a lot about procedural content generation, and where automated game design fits into it. Since I started work on ANGELINA I’ve kind of always lumped it in with all other procedural content research, but the more I think and talk about the system, the more distinct automated game design becomes as a term. Here’s a few hundred words on the idea, how the two differ, and why it might become important in the future.
The reason I began thinking about all of this is because
I’ll be crediting ANGELINA as the chief (and possibly sole) level designer in A Puzzling Present, and it occurred to me that this is the first time in videogame history that an AI has been credited as such (of course I may be forgetting things here – do let me know below!)
Update: I’m no longer crediting ANGELINA/Mechanic Miner as the sole level designer on A Puzzling Present, partly because of computational creativity questions raised by some people I’ve talked to since, and also because I’ll be curating the levels. It’s certainly an issue for the future though – I think the rest of this post is still worth discussing!
I was sure that this was true, but there are lots of games that have had their levels designed by AI techniques, so why did it feel like there was a distinction here?
Really, it’s because the motivation behind ANGELINA is to build a system that produces static content, designed in the much the way a human might, and released as a single consumable item. By contrast, most PCG systems are designed to produce content either ad infinitum (for players to consume endlessly) or at the very least in large quantities during pre-production (from which a human designer selects or tweaks certain chosen content). These approaches are at odds with each other in many ways.
For instance, many new procedural generation techniques are sold on their ability to generate content quickly, or to terminate early and deliver content whenever the player needs it. These are obviously useful for generating content at runtime, but for automated game design? We often don’t care. It doesn’t matter if ANGELINA takes an hour, a day or a week to produce a game, because like human developers ANGELINA can take a ‘done when it’s done’ approach and take its time to produce a design that it is happy with publishing.
Similarly, we’re seeing a lot of adaptive PCG in research at the moment, where content is not only generated at run-time, but is generated to suit the player in some personal way. This is an explicit anti-goal for my work with ANGELINA – it’s actually more important to me that the system is able to think about the impact its content will have on groups of people, particular cultures or ways of thinking, rather than individuals. ANGELINA produces content to be consumed once, and shared by everyone.
This leads to some weird asymmetry between the two. Whereas PCG might be designed with a human fitness function in mind (such as experience-driven PCG or mixed-initiative tools designed for a game designer) AGD has to explicitly avoid this, because there are no humans involved in the design process (until we start looking at AI systems that crowd-source human playtesters during their execution). Thus, techniques that work well in PCG may be unusable in AGD, and AGD techniques may be so reliant on automation that they don’t make sense in PCG.
I still feel that AGD is contained within PCG, but it feels like a very well-defined subfield that doesn’t fit all the ways in which we understand PCG to work up until now. That’s exciting to me because it means we have to think about things a little differently, but I think it might also mean that we need to be very careful when reviewing or comparing AGD work to make sure that we’re not looking at it as we might standard PCG. The ways in which the two approaches differ are quite meaningful, and over time will probably diverge entirely. I’d love to hear what other people think about this, whether I’m separating two identical areas, or whether I’m right to draw a distinction between the two.
Julian @togelius and Mark Nelson spoke a little about this with me on Twitter, with Julian raising the idea that perhaps this is just offline PCG, rather than online. It’s an interesting point that made me think for a bit. I’m not sure if it totally covers my way of thinking (offline PCG implies a human curator, whereas AGD to me says that we are pushing for autonomy – so ANGELINA’s levels designed for A Puzzling Present are not curated in any way, they are used as-is by the system) but it’s a good point. Offline and Online PCG are described, along with other types of PCG, in this great survey paper by Togelius et al.