{"id":8443,"date":"2018-03-07T14:48:33","date_gmt":"2018-03-07T14:48:33","guid":{"rendered":"http:\/\/rafaelfajardo.com\/portfolio\/so-you-want-to-build-a-generator\/"},"modified":"2018-03-07T14:48:33","modified_gmt":"2018-03-07T14:48:33","slug":"so-you-want-to-build-a-generator","status":"publish","type":"post","link":"https:\/\/rafaelfajardo.com\/portfolio\/so-you-want-to-build-a-generator\/","title":{"rendered":"So you want to build a generator…"},"content":{"rendered":"

galaxykate0<\/a>:<\/p>\n

\n

\"image\"<\/figure>\n<\/p>\n

This is a beginner-level advice essay for people just getting started with building generators. It\u2019s also for practiced experts who want a way to organize their knowledge. The advice is meant for any kind of generators: humorous twitterbots, fanciful art-making bots, level generators, planet builders, makers of music, architecture, poetry, and cocktails. (edit: it turned out to be 4000 words long, so enjoy the ride!)<\/p>\n

With so many possible kinds of generators, what advice can I give? Any advice will depend on what kind<\/em> of generator you want to build. I\u2019ll give you a list of questions<\/strong> to ask yourself, and advice to match the possible answers.<\/p>\n

The first question is:
\nWhat are you making?<\/strong><\/p>\n

Write down the thing<\/em> that your generator will make. We\u2019ll refer to this as your \u201cartifact\u201d but it could be anything: procedural birds, generated stories, animated dance choreography, gazpacho recipes, RPG quests, chess variants.<\/p>\n

Now the hard part. Inhale, exhale, and start writing down everything you know about what makes one of those successful. What qualities does a good one of those artifacts<\/em> have? \u201cFunny\u201d, \u201cHarmonic\u201d, \u201cPlayable\u201d? Now go deeper: the more specific the kind of thing you are making, the more specific and detailed you can get with these requirements. What does a \u201cfun\u201d level mean? \u201cFun\u201d will mean very different things for Super Mario or Civilization or Bejeweled. I can come up with more properties of a good paranormal regency romance novel<\/em> than I can come up with rules for a good novel<\/em>. The easiest generators to make are the ones where you can describe \u201cgood\u201d artifacts as sets concrete properties.<\/strong><\/p>\n

Now flip the question on its head: what makes a \u201cbad\u201d example of this artifact? List everything you can think of, including properties that are merely unfortunate, and properties that are total deal-breakers. Put a star next to anything that absolutely must not happen<\/em>. These are your constraints<\/em>. The most reliable generators are the ones where you can concretely describe constraints<\/strong>.<\/p>\n

You now have a list of desirable properties<\/strong> and constraints<\/strong> that you want and don\u2019t want for your artifacts. We want a generator with a possibility space<\/em> (all the kinds of artifacts it can generate) where most of the artifacts have the good properties, and few (or none) of the artifacts have bad properties. We\u2019ll also want a range<\/em> of artifacts, not just the same perfect artifact over and over and over, though how wide<\/em> a range that is is a choice for you, the designer (a problem identified by Gillian Smith<\/a> and being explored by Michael Cook<\/a>).<\/p>\n

Now we have the guidance we need to start building methods<\/strong> that make artifacts.<\/p>\n

An alternative approach: the previously described method is good for inflexible situations, where you know ahead of time what you want to build. For many situations, like game jams or prototypes or side projects, you can be more flexible and improvisational! You can start with a methods or some loose pieces to figure out what it \u2018wants\u2019 to generate (what properties is it best at generating?) and then revising what artifacts you are generating to better suit what the generator is good at making. I talk more about this approach in my 2015 Procjam talk<\/a>.<\/p>\n

\"image\"<\/figure>\n<\/p>\n

Building Your Artist-in-a-Box<\/h2>\n

When the Spore editors were being designed, the engineers, designers worked closely with the art team to understand how<\/em> an artist or an animator would go about sculpting and texturing a creature. If they could understand the process<\/em> and design an algorithm that could follow that process on demand, they would have an \u201cArtist-in-a-box\u201d that could make procedurally generated creatures that would be nearly as good as the original animator would have made. Designer Chaim Gingold explained this process in his 2007 GDC talk audio<\/a> slides<\/a> and Art Director Ocean Quigley used a similar process to experiment with what kinds of building blocks he would want for building cities<\/a><\/p>\n

It is helpful when making a generator to sit down with someone who makes the sort of artifacts you are building, and have them walk you through the process of making something. What questions do they ask themselves along the way? How do they make decisions? How do they describe the tradeoffs between choices? How do they describe the different problems that they have to keep in mind? How do they name all the parts of what they are working on, and all the relationships between them? (the academic term for this is an \u201contology\u201d)<\/em><\/p>\n

Some fields have expert practitioners who have written frameworks<\/em>, often controversial, to describe what they do. Music theory has proposed plenty of rule systems, for example, for jazz improvisation or Bach-style harmonies or pop songs. Story generation has narratological theories like The Hero\u2019s Journey, but also informal frameworks like TV Tropes. Art theory has the Golden Ratio and color harmonies and composition rules (I haven\u2019t found those visual aesthetic rules to be productive in my own work, though you may feel differently). No framework is complete, or makes provable good artifacts, but they can help give you guidance and inspiration.<\/p>\n

So now, ask yourself: How would a human solve this problem?<\/strong><\/p>\n

\"image\"<\/figure>\n<\/p>\n

From rules to generative methods<\/h2>\n

Unfortunately, knowing how a human would make something isn\u2019t the same as being able to teach a computer how to do it. Humans are good at estimation, and making guesses, and synthesizing a lot of knowledge about past situations. Computers know only what you tell them, and many problems require way more implicit knowledge than we think, but they are<\/em> good at performing lots of calculations and trying out lots of possibilities. So the methods we want to use will need to provide a way for the computer to solve problems like<\/em> a human, or at least with a way to mirror some of the skills of a human. Methods that are particularly good for building generators (generative methods<\/strong>) will give the computer some of the following skills:<\/p>\n