All art below was generated by interactive art evolution programs incorporated into
MM-NEAT, a software
package that evolves artificial neural networks. Each piece of art presented below was
generated by a special type of neural network known as a Compositional Pattern Producing Network (CPPN).
Specifically, human subjects used interactive evolution to select the art they enjoyed the most,
resulting in the generation of more art in line with their preferences. Although the larger MM-NEAT
project still retains the ability to evolve such artifacts, a simplified version of the code that
focuses exclusively on interactive evolution with CPPNs is also available at
https://github.com/schrum2/CPPNArtEvolution.
This code was written by Isabel Tweraser and Lauren Gillespie as part of Southwestern University's
SCOPE summer research program.
Art Selected By Human Subjects
Selecting the name of one of the art generation programs on the left will provide you with a list
of individuals that participated in a human subject study to understand the capabilities of these
systems. Users were either exposed to both Picbreeder and AnimationBreeder, or 3DObjectBreeder and
3DAnimationBreeder. Therefore, subject ID numbers for the two 2D art programs correspond to the same
person, and subject ID numbers for the two 3D art programs also correspond to the same person (for a total
of 40 participants). Therefore, you can compare what one individual generated with one program to what that
same individual generated with another program. However, no subject used both 2D and 3D art programs.
Users interacted with each program for 15 generations. In each generation, 20 options were available
to select from, but only the items that each user actually selected are accessible in the interface below.
Images and animations are reduced in size for the sake of space, but the original CPPNs can represent
the art depicted here at arbitrarily large resolutions.
This remake of Picbreeder allows a single user to evolve artistic images using Compositional Pattern Producing Networks,
just like the original. However, the interface has some extra features offering more control over the process.
AnimationBreeder
The AnimationBreeder adds a time input to the evolved Compositional Pattern Producing Networks so that multiple images like those produced by Picbreeder can be produced by a single CPPN as a function of time.
3DObjectBreeder
The 3DObjectBreeder uses Compositional Pattern Producing Networks to evolve 3D shapes using the same interface as Picbreeder and AnimationBreeder.
This program is a remake of EndlessForms, though it does not use the Marching Cubes algorithm to smooth out the evolved shapes.
However, it does allow the color of each voxel to be evolved, and also allows for slight displacements in the location of each voxel so that strict adherence
to a grid of voxels is not required.
3DAnimationBreeder
The 3DAnimationBreeder adds a time input to the Compositional Pattern Producing Networks from 3DObjectBreeder in order to animate the
3D shapes that are formed. The presence, color, and displacement of voxels can change over time.
Jacob Schrum,
Pier Luca Lanzi,
Alexander Nareyek,
and Pieter Spronck
(2018).
Human-assisted Creation of Content Within Games,
Dagstuhl Reports: Artificial and Computational Intelligence in Games: AI-Driven Game Design (Dagstuhl Seminar 17471), Volume 7, No. 11, page 111. Editors: Pieter Spronck and Elisabeth André and Michael Cook and Mike Preuß. Schloss Dagstuhl--Leibniz-Zentrum fuer Informatik
Fall
2018:
The cover of SIGEVOlution Volume 11, Issue 4 features art generated by AnimationBreeder, the interactive evolution system described in a GECCO 2018 paper co-authored with SU students. SIGEVO is the ACM Special Interest Group on Genetic and Evolutionary Computation.
Summer
2018:
Neuroevolution in Video Games: "Mad Science Monday" presentation
made by my SCOPE Summer research students to present to other SCOPE students