| by Tatsuo Unemi, Soka University. August - October 2016. | Tweet |
|
Selections by Artificial Critic #1 from Tatsuo Unemi on Vimeo. |
A computer-based installation that shows the Selections by the computer from the five years’ stock of Daily Evolved Animations. Keywords: evolutionary art, evolutionary critic, artificial critic. |
| As a challenge to make the machine that makes art, a machine in the author’s laboratory has been working to create ten art pieces everyday since October 2011 under the project named Daily Evolved Animations. Each of the pieces is a type of animation of 20 seconds composed with abstract drawings as the frame images, evolutionarily selected based on a computational aesthetic measures through 300 steps of generational changes. More than 18,000 pieces in total have already been stocked in the server machine so far. Though the author carefully designed the evaluation criterion for the genetic algorithm implemented in the system, the results still do not always satisfy the human’s aesthetics. It would be a reasonable strategy to improve the fitness criteria as to fit more with human’s aesthetic measure, but it is also nice to collect the better pieces from the stock according to evaluation by a type of Artificial Critic in order to save the five years of efforts by the system. To design an appropriate evaluation function, the author choose a number of pieces from 920 pieces produced from January 1st to April 1st of this year, as both positive and negative examples for training a machine-learning system. The target function is to calculate the grade point for each piece based on the values of the twelve elemental features used for evolutionary system of daily production. |
We already have several types of methods to solve such type of optimization problem such as Statistical Cluster Analysis, Artificial Neural Networks, Support Vector Machine, and so on. Here the author employed a technique of Genetic Programming as the first trial. Through thousands of steps for each of a number of different settings of the algorithm, it found a function that can grade all of 920 samples not perfectly but in acceptable level.
This installation is to display the selections from all of the pieces in the stock from October 14, 2011 to September 2, 2016. The selections contain 3,560 pieces that got higher grade by the evaluation function found though the above algorithm. Because the total duration for all of these pieces is almost 20 hours, the installation will show the pieces in turn within the allowed hours in the exhibition. The data for these animations are not in a form of movie file but program fragments in a shading language of OpenGL. This feature is useful not only to reduce the size of data storage but also to realize lossless images for each frame in high resolution even in 4K. It will be displayed using a 4K monitor or a Full HD projector depending on the availability of the equipment and the environmental situation of the exhibition site. |
 Evolutionary process to find an optimal function. |
 Lists of pieces with feature values. The third column indexes G (good), N (nuetral), or B (bad) by the author's subjective evaluation. Left side is in the chronological order of original production. Right side is in the order sorted by evolved function. |