Notes on Practical Procedural Generation

procedural-generation
notes
My notes from Kate Compton’s talk on practical techniques for procedural generation.
Author

Christian Mills

Published

December 29, 2021

Overview

Here are some notes I took while watching Kate Compton’s talk covering practical procedural generation techniques.

Examples

Minecraft Official Site

No Man’s Sky

Bay 12 Games: Dwarf Fortress

These Monsters

Cameron’s Yavalath Page

  • Made a system that could generate game rules
  • Made a player that could play arbitrary games
  • Had virtual players play thousands of games, until he found a game that was pretty well balanced

PANORAMICAL on Steam

It is as if you were playing chess

Fitzwilliam Darcy’s Dance Challenge

The Treachery of Sanctuary - CHRIS MILK

Kinematics Dress

V&A Design a Wig

Toca Hair Salon - The Power of Play - Toca Boca

  • Lots of generative content uses extremely sophisticated and brilliant AI and fails anyway
  • Some of the best generative content is simple
  • The hardest part of procedural content is design

Steps

  1. Understand the design space
  2. Enumerate your constraints
  3. Understand the process
  4. Pick a generative method
  5. Iterate and be flexible
    • A lot of great generative projects are things that were tried because it is a stupid idea

What are you making?

  • Be specific
    • Level generator
    • Character creator
    • Abstract art generator
    • cocktail recipe generator
    • game title generator
    • conversational character
    • poetry generator
    • twitterbot

Making an artist-in-a-box

  • teaching an algorithm to make art like an artist
  • Find and expert (or read their writing)
    • How do they think through a problem?
    • Example Question: “If you are designing a creature, what do you do?”
    • Example Answer: They start by drawing a bean shape as a base for the creature, and hangs a mouth on it.

Additive and Subtractive Methods

  • Build up a space of good stuff
  • (optional) Remove bad stuff
  • Vocab:
    • Possibility space
    • Expressive range

The IKEA Catalog of Generativity

  • A catalog of generative methods and why you might chose each

Additive Methods

Tiles

  • Works well for
    • Something you can break into (equal-sized) regions
    • where tile-to-tile placement don’t need to be constrained
      • Can use WaveFunctionCollapse when placement needs to be constrained
    • but you can still get emergence from the placement of tiles
    • one of the oldest forms

Grammars

Distribution

  • put down a bunch of stuff
  • can use random numbers (actual randomness does not look good)
  • real distributions are hierarchical and clustered, but also maintain spacing
  • Barnacling: when you have a large object in your world, there should be medium sized objects around it and smaller objects around those
  • Footing: When two things intersect, there should be an awareness of them intersecting
    • Example: If you stick tree in the ground, there will be dirt piled up around it
  • Greebling: cosmetic detailing added to the surface of an larger object that makes it appear more complex or technologically advanced
  • Options
    • start with a grid, and offset a bit
      • (less obvious with a hex grid)
    • Use a voronoi diagram with easing
    • Do it properly with a Halton Sequence

Parametric

  • An array of floats representing settings, “morph handles”
  • modellable as points in an N-dimensional cube
  • Any position is a valid artifact
  • You can do genetic algorithms
    • or use directed walks through the space
    • or “regionize” the space

Interpretive

  • Start with an input
    • Run an algorithm to process data into some other data
  • You have a simple structure
    • some distribution of points, a skeleton, a connectivity map, a curve or path and want to make it more complex
  • Examples:
    • Noise (Perlin/simplex)
    • Voronoi/Delaunay
    • Constructive Solid Geometry Extrusion, revolution
    • Metaballs
    • Fractals, mathematical models of impossible shapes
    • (Hypernom, Miegakure)
      • low control, high weirdness, not suitable for most games

Simulations

  • Particle trails
    • simulate particle path responding to forces
  • draw directly
  • OR record path and use for extrusions or distributions (Photoshop brushes)
  • Goes great with user input (Leapmotion, Kinect)
  • Cellular automata
  • Agent-based simulations
  • Physics simulation

Subtractive Methods

Saving Seeds

  • Seeded random numbers
    • Same seed, same random generation
      • Make sure nothing is framerate or input dependent
  • Whitelist a catalog of known good content
    • It’s faster to verify questionable content than to build a testing function

Generate and test

  • If you can write an algorithm to judge “quality”
    • Throwaway vs ranking/prioritization
      • Use ranking/prioritization
    • Test for brokenness/connectivity
  • Beware of false functions
    • beware the “fun equation”

Computationally exploring the possibility space

  • Also called “search”
    • Brute force search
      • “Find the tallest creature that the tool can make”
      • “Make a level that has these properties”
    • Hill-climbing
      • Genetic algorithms
      • Works best with parametric methods

Constraint-solving

  • You can describe a possibility space and constraints, just find the valid parameters.
  • Inverse Kinematics-solving
  • Answer set solving
  • Brute force
    • pay attention to exponential growth

Making use of Generativity

  • You can generate many things
  • They are all mathematically unique
  • But they aren’t perceived as unique
  • Is this a problem?
    • Do not boast about really big numbers

Different kinds of generative content

  • Background
    • In-fill (don’t be empty)
  • Perceptual differentiation
  • Perceptual uniqueness
  • Characterful
    • Test: Would you write a fanfic for this generated item?

Ownership: MSG for PCG

  • Allow users to name content
  • Showing off content with their name attached, to a large audience
    • The “victoriain explorers club” model
  • promote players
  • Let players take credit for your generativity
    • creators, curators, retellers

Data Structures: Make your life easier

  • A/B test generators
  • Release new generative content safely
  • Create editors and run user-made generators safely
  • Visualize your generators

Further Reading

References:

About Me:

I’m Christian Mills, a deep learning consultant specializing in practical AI implementations. I help clients leverage cutting-edge AI technologies to solve real-world problems.

Interested in working together? Fill out my Quick AI Project Assessment form or learn more about me.