(modified Oct 30, 2024)

The Buffalo AKG Electric Op exhibit reminded me of an old folly, and this is a second iteration. Briefly, I am using a simple population genetics model of genetic drift with infinite alleles as an excuse to learn a bit more JavaScript, including the p5 library.

Bubbles

The previous iteration used an individual-based simulation to follow a population of individuals over time. The innovation here is to follow the small number of segregating alleles, rather than the 1000’s of individuals in the population. The simulation is both smaller and faster, so more realistic population sizes can be used. The simulation may still have some artistic merit…

The population size N is 10 times larger, and the mutation rate μ 10 times smaller, than in the individual-based simulation. The number of generations per frame is increased four-fold to 200, and the frame rate doubled to 24 (200 x 24 = 4800 generations) per second (available CPU may limit the rate of display; this animation is slow on mobile devices). As before, ‘Segregating’ alleles are just the number of alleles in the population at each generation. ‘Replacements’ summarize how many times a common allele has been replaced by a new allele.

Notes on representation

Although the figure is ‘mathematical’ and in that sense contrasts with Mondrian’s carefully composed paintings, it is useful to reflect on decisions made during development.

The bubble chart representation is much less frenetic than the grid used in the individual-based simulation (I could have summarized the individual-based simulations using bubble charts, but then the connection to Mondrian would be even less clear).

Originally I set the maximum bubble diameter to 1/2 the minimum width or height of the canvas. This meant that new alleles were centered in a two-dimensional square on the canvas. The simulation above sets the maximum diameter to the minimum width or height, so when only a single allele is present its bubble occupies the entire width or height of the canvas. A corollary is that new alleles are constrained to a vertical (if the width of the canvas is smaller than the height, as on a mobile device) or horizontal (width larger than height, as in a browser) line. This further simplifies the bubble chart dynamics, with the eye needing to cope with fewer changes across frames.

The background color was originally white, but switching to black is easier on the eyes and seems to better contrast, in general, with the randomly chosen allele colors.

The alleles were originally drawn without an explicit stroke for the circumference. Drawing the circumference provides a more defined shape for the eye to perceive. This is particularly important with partially transparent bubbles, and transparent bubbles are needed to visualize overlapping alleles.

Implementation notes

In addition to p5, I use binomial() from stdlib. Here’s the Boogie Woogie bubble JavaScript.

The script is written using p5’s ‘instance’ mode, and as a JavaScript module. Using instance mode and modules loads only p5 variables into the global namespace, not stdlib symbols or symbols used in the Boogie Woogie script. Instance mode involves writing a closure (function) that is passed to the p5() constructor. Simulation parameters and functions are defined (currently) in the closure rather than at the level of the module, resulting in better encapsulation. The stdlib symbol binomial is imported into the module, more closely coupling the import to its use. Modules also allow more careful validation, flagging a couple of instances where variables were used without let or const declarations. Using p5 instances also allows for more than one canvas on a page.