Mu Cartographer is a game that wouldn’t exist without generativity.
As you explore the fluctuating landscape, uncovering snatches of story, you quickly realize that this is something different than your usual map-exploration game. It doesn’t just generate a map: it generates an entire multidimensional space within the manifold of its mathematical model, and then asks you to explore it.
Learning how to operate the alien machine that lets you navigate this world is part of the fun, so I won’t go into too many details about how it works. But it does work, giving surprisingly effective hints as you tease out the principles that govern its inner workings.
This is a very unexpected use of the kind of flexibility that isn’t possible without procedural generation and related techniques. I like it. It gets at the exploration of abstract interactive systems that I want to see more of, and it visualizes those unseen spaces in an accessible way.
Keep your grad students from stressing out! Watch them brainstorm! Double click on ideas on the whiteboard to start projects! Try to get them to finish writing papers in a timely fashion! Submit them to conferences and get terrible feedback from Reviewer Two!
There’s a ton of Tracery-powered procedural content generation going on here. I especially like the paper titles, and some of the stuff that gets written on the whiteboard describes projects I’d want to work on: Long-term Strategy + Tabletop RPG = ? Maps + Metroidvania->??? LudoForm: clustering for MetroidVania maps.
Automatic Generation of Typographic Font from a Small Font
Subset
Which is a mouthful. But pretty neat. This research by
Tomo Miyazaki, Tatsunori Tsuchiya, Yoshihiro Sugaya, Shinichiro Omachi, Masakazu Iwamura, Seiichi Uchida, and Koichi Kise
is exactly the kind of centaur-aid design tool that I want to see procedural generation used for.
Instead of designing a font by hand, the designers can design a few and generalize to many more. Which is important if you want to speed up the design of kanji characters. As they mention in the paper, with thousands of characters to create, a professional designer can take years to make a single font for CJK languages.
Interestingly, they’re using images as source data rather than vectors, which are more usual. They say they chose to do this because it’s much easier to acquire image data for new characters, whereas vector data usually needs to be constructed.
It still needs a vector skeleton to work off of, so totally novel characters need to be annotated manually. Though they’re using the GlyphWiki dataset of 270,000 characters as a base, so they may have the character you’re looking for already.
Bots like this are a very dadaist form of protest against the absurdity of the world.
The original Dada movement was reacting to the carnage of the first World War, by rejecting the systems that lead to it.
While they were deeply skeptical about technological progress, the Dada artists also innovated by using odd materials and randomness to make anti-art.
The processes of the bot reemphasize the original message, highlighting the absurd truth by contrasting it with similar examples.
Made by
Jacob Garbe and running on CheapBotsDoneQuick, it’s not quite Dada, but then again Dada isn’t Dada.
BotRoss is a Processing-based twitterbot that posts generated images like Bob Ross would have programmed in 1985. Not just paintings in the style of, but following the processes he used.
The conceit is that the generator was found on a floppy disk in a PBS station, but while I often go along with the author’s framing, I want to pull back the curtain a bit here and highlight some of the technical details here.
Specifically, the bot doesn’t use Perlin Noise. This is partially because Perlin noise wouldn’t have been available in 1985, but also because it gets used a lot in other generators. By making the artistic choice to avoid it, it has the additional benefit of giving the generator a unique look.
The other artistic choices (the dithering, following Bob’s process closely) similarly help both the framing and the aesthetic output. Not to mention, it generates painting that would be at home on a C64 RPG.
“Pyramid of Arclength Descriptor for Generating Collage of Shapes,” by Kin Chung Kwan, Lok Tsun Sinn, Chu Han, Tien-Tsin Wong, and Chi-Wing Fu,
is a research paper on filling a specified shape with a collage of images.
The results are quite nice:
The technique supports a number of additional settings and constraints. They can even loop infinitely:
While the authors concentrate on the use in collage and design (and this will undoubtedly be a highly-useful design tool that someone will rip off for a music video or ad in the next year or two) I can’t help think of the wider applications.
As I’ve said before, any procedural generation technique can be attached to many other techniques. What if you built a map generator out of this? It wouldn’t need to be constrained to rectangular shapes; you could use any curving shape.
Or, more mundanely, laying out a texture sheet or a UV atlas. While there’s been quite a lot of work in this area, existing algorithms have their own drawbacks. Broader still, packingproblems are very common in many industries. How many shapes can you cut out of one board? How do you load the most items on a pallet?
While this algorithm doesn’t guarantee the maximum efficiency, it does work with arbitrary irregular shapes, and it uses a search space rather than a time-intensive process like physical simulation.
A talk at VideoBrains by Tommy Thompson about procedural generation and Borderlands 2.
Starting off with the connection between the rate of interaction, the prominence of the procedural generation, and the variety in the system, he draws a connection to looking for a gun that you love in Borderlands.
He argues that the experience of trying to find the outlier in the the data, the gun that’s perfect for your current point on the level curve, is the greatest strength of the game, rather than the details of the procedural generation system.
There are a number of features that make the maps in SMAC more interesting than those in the earlier Civilization games. It uses height to make hills and mountains, for example, instead of them being tile-features. The special resources are distributed in more interesting patterns; the newly-introduced borders make the size of the map work better; and the native lifeforms are better integrated that the barbarians were (or are, for that matter).
But the map generator has two really interesting features that still set it apart from other Civ-style games.
The first isn’t a feature of the generator, per se, but greatly affects the meaning of the maps: the player can terraform the planet. And not just in little ways, like raising or lowering a couple of tiles, though you can do that too. A couple of the council resolutions can raise or lower the sea level across the entire planet. (Global warming from too many boreholes can also melt the ice caps to the same effect.) The malleability of the terrain makes it fairly unique among strategy games.
It can be a viable strategy to flood the map and drown your opponents cities, or to drain the ocean and march your armies across on dry ground.
The second is vital part of the generator: the landmarks.
When a map is generated, it scatters a number of prefabbed features on the planet. A few are mostly decorative, but most have an effect of some kind.
They owe a bit, I think, to the discoveries in Seven Cities of Gold (like the Grand Canyon) and the wonders in Civilization–the manual refers to them as “giant natural wonders of Planet”–possibly via Colonization, though at the moment I can’t remember if that game had any exploration bonuses for natural wonders.
The landmarks in Alpha Centauri are unique even when compared to the later Civ games that included similar features. They occupy multiple map tiles, sometimes forming significant strategic features on the map in addition to their resource bonuses.
Moreover, they help give the random maps structure. In contrast to the accidental chokepoints of earlier Civ maps, they have deliberate strategic importance.
The map generator as a whole is “spikier” than earlier random Civ maps. The landmarks make things a bit less fair but more interesting. There were a lot of high-value city-sites in Civ II because the even pattern ensured that they would be frequent and predictable, but there’s only one Manifold Nexus.
Which is not to say that it’s an absolutely dominant strategy: There’s enough landmarks overall that everyone should be able to claim one, if they work on it. But there’s plenty of other things going on, so you may have other priorities.
In the end, it is a good demonstration of how maps (and procedural generation in general) are much more interesting when they have outliers to act as landmarks and memorable setpieces.
The really interesting thing about the guns in the Borderlands series is that they aren’t completely random.
Their procedural generation has a structure. Moreover, that structure is reflected visibly in the 3d models and invisibly in the properties of the weapon. You can, if you’re familiar with the system, tell what properties a gun has by looking at it.
Each weapon is assembled out of a set of parts, has a manufacturer (each with unique effects), a title (selected by the combination of the the body and barrel), and so on. Each manufacturer has weapon types that they don’t produce, breaking the symmetry of the pattern.
The manufacturers, by the way, are a great way to introduce lore to your procedurally generated items. It gives a really strong flavor to the different categories of the generator and gives the player a narrative hook that you can hang a mechanical effect on.
The recently finished concert hall uses algorithmically-designed paneling (via the computational geometry of 1:One, who specialize in similar architecture and sculpture):
Using these requirements as parameters, Koren developed an algorithm that produced 10,000 panels, each with a unique shape and pattern, mapped to clear aesthetic and acoustic specifications. “That’s the power of parametric design,” he says. “Once all of that is in place, I hit play and it creates a million cells, all different and all based on these parameters. I have 100 percent control over setting up the algorithm, and then I have no more control.”
Which reminds me a bit of the generation of the Deep Note, and in a more general sense, the use of procedural generation to create designs that are too complex for humans to create or describe by hand.
It also suggests that we’re entering an era where design is too complex for humans to easily describe. Historically, the mathematical models used in architecture were complex but were able to be reduced to simple algorithms. For example,
Vitruvius’s description of the proportions of a column.
But, similar to computer-assisted proofs in mathematics, we can now design things with properties we don’t fully understand. No human knows the exact effect of each panel in the concert hall. No person has a complete model of the results in their head.
Of course, we could still reverse engineer it, treating it as an enormous puzzle. But that just emphasizes that our centaur design methods are now capable of creating results more complex than humans can imagine in detail.
