Aesthetics and the Edge of Chaos

Transcription

1 Aesthetics and the Edge of Chaos Dora Raymaker 2008 SYSC557 Artificial Life I wasn't able to get James Earl Jones to narrate for me, so you'll have to be con tent with my computer doing it. If you can hold your questions until the end, I would very much appreciate it. This is my project on aesthetics and the edge of chaos. I'm a systems and computer science graduate student, but my undergraduate degree is in fine arts, which may explain why I'm drawn to the intersection of artificial life and fine art.

2 Background: Art Modeling Nature Art reflects nature Automatons Visual arts Pollock's fractals Evolutionary reason for aesthetics? Attracted to things that improve survival Too much order = dead? Too much chaos = dangerous? Jackson Pollock Blue Poles: Number II, detail Art is one of the oldest methods people have used in modeling nature. Langton touched on this in his discussion of auto ma tons and water clocks in our initial class readings. But art modeling nature isn't restricted to three dimensional, robot like art. Paintings, drawings, prints, photography, and other forms of visual art can reflect nature and life. Even highly abstract work, such as Pollock's blue poles, was found to have a fractal structure similar to patterns found in nature. Some psychologists believe there may even be evolutionary reasons for aesthetic preference, in that we are drawn to things which would improve our chance of survival if they were encountered in reality. As things in nature which are very orderly or very chaotic tend to either be dead or pose a threat, would this imply people would find edge of chaos more aesthetically appealing?

3 Background: 2-D Design Principles Basic 2-D Design Tension of order and chaos repeated patterns and established structure novel surprises and lack of established structures Tension needed to bring the page alive Kandinsky illustrates the use of points and lines in the Appendix of his text Point and Line to Plane. This describes Edge of Chaos? One of the basic principles taught in art school is that tensions between ordered and chaotic elements makes an image come alive. I recall an exercise which involved putting a black dot in various places on a white square. If the dot is placed in an orderly location, such as at the center of the square, the image looks "dead." If however the dot is placed slightly off-center, not quite predictable, the entire white square activates visually and comes "alive." Too many repeated patterns and established structures, and an image is considered boring and dead. But too many novel surprises and lack of structure, and the image becomes boring and mud. Images which balance repetition and novelty are interesting and exciting, the tension between them is what makes a composition alive. The artist Kandinsky wrote a lot about this dynamic. Perhaps unsurprisingly, excerpts of that writing, including the image on this slide, appear in a paper published in artificial life. This was a main motivation for this project. I have always been struck by how edge of chaos dynamics echo the order chaos tension that is considered necessary for bring ing "life" to a work of visual art. Does this principle and edge of chaos describe the same thing, just in different fields?

4 Background: Creative Process Creativity from tension between Analytic thinking (structure) Associative thinking (novelty) Illustrated at right :-) Edge of Chaos dynamics in creative process itself? Dora Raymaker Machiname Order chaos tensions seem to exist on all levels of both life and art. Some theories of creativity posit that the creative process itself is a tension between ordered analytical thinking and chaotic associative thinking. Analytical thought provides constraints and structural organization, while associative thought provides "thinking outside the box" and endless novelty. The image on this slide is one of mine, and it was an illustration of my own creative process. I made this painting before I'd known about this theory of creativity, but it is actually, quite explicitly, about analytical versus associative tensions. Is there a need for both orderly and chaotic, edge of chaos, type dynamics to bring the creative process itself alive.

5 Background: Art and Artificial Life Paul Brown: The Deluge - after Leonardo, Giclée Print, 50 x 50 cm, 1995 Two time steps are combined and mapped onto a set of four tiles. Here the tiles represent transitions in the underlying automaton rather than the states themselves. Artists embrace Alife e.g. Paul Brown ALife researchers embrace Art e.g. Karl Sims Karl Sims: Genetic Images is a media installation in which visitors can interactively "evolve" abstract still images. A supercomputer generates and displays 16 images on an arc of screens. Visitors stand on sensors in front of the most aesthetically pleasing images to select which ones will survive and reproduce to make the next generation. Maybe because of these strong relationships between art and life, scientists working in artificial life and artists embraced each other's ideas early on. There is less conflict between "art vs. science" when it comes to artificial life, and more of an attempt to unify the two and work cooperatively. From the art side, artists such as Paul Brown have been using artificial life techniques in their work even before Langton coined the term. Brown first started drawing cellular automata by hand without use of a computer in the 70's. The work pictured on the left is a C ay in which two time steps are combined and mapped onto four tiles, so the tiles represent the transitions rather than the states. From the artificial life side, scientists such as Karl Sims have used artificial life both as a means to generate art, and art as a means to explore questions in science. The image on the right is one generated through an interactive process which evolves still images. A computer genner ates and displays images on an arc of screens, and humans stand on sensors in front of the most aesthetically appealing images. The humans act as the fitness function for the gee ay which evolves the next batch of images.

6 Hypothesis People will find images which exhibit Edge of Chaos dynamics more aesthetically appealing than images with other types of dynamics. Because of the relationship between art and nature, the order chaos tensions of design principles, the nature of the creative process, and the natural affinity between art and artificial life, I became interested in exploring the relationship between chaosity and aesthetic appeal. My hypothesis is that people will find images which exhibit edge of chaos dynamics more aesthetically appealing than images with other types of dynamics.

7 Definitions Aesthetically appealing in 2-D visual art will make one crazy trying to rigorously quantify! in the eye of the beholder beautiful, interesting, or appealing based on first impression or gut reaction assured this is OK to do in social science research! Dynamic regimes Fixed Point (repetition of same state) Limit Cycle (repetition of sequence of states) Edge of Chaos (some repetitions / some novelty) Chaos (continuously novel states) So the first tricky step in working with this hypothesis is defining what, exactly, aesthetic appeal is. First I defined the domain as 2 dimensional visual art. Then I made myself insane trying to figure out how to quantify the idea of aesthetic appeal. Luckily I have friends who do social science research, and they steered me away from the crazy making, telling me that it's O K to allow people to define subjective terms for themselves. So I ended up just defining aesthetic appeal as beautiful, interesting, or appealing, and based on first impression or gut reaction. I then required additional assurances that it was O K to do this. Dynamics on the other hand are much less squishy. The dynamics are defined by attractor type, and are fixed point, limit cycle, edge of chaos, and chaos.

8 Variables and Confounders Variables of interest Ranking of aesthetic appeal Dynamic regime of image Complexity of image Possible confounders Visual elements: colors, shapes, textures Image properties: size, scale, proximity, and sequence Participant fatigue The variables of interest in this study were ranking of aesthetic appeal, the dynamic regime of an image, and the complexity of an image. Possible confounders that would influence someone's perception of aesthetic appeal included all visual design elements such as colors, shapes, and textures, all global image properties such as size, scale, proximity, and sequence, and additional perceptual and psychological factors such as viewer fatigue. In some of the next slides I'll address strategies for accessing the variables and avoiding the confounders.

9 Design Overview Part I: Image Generation Part II: Survey Part III: Results (analysis) The general study design consisted of three sequential parts. The first part considered image generation, how to create interesting images with unambiguously quantifiable dynamics. The second part was a survey, given to human participants, which asked them to rank the aesthetic appeal of the images generated in the first part. The third part was to analyze the results of the survey in some meaningful way.

10 Image Generation: Technical Considerations Cellular automata Pros Visualize dynamic regimes Classifiable Visually interesting Learning opportunity K = 4, N = 5, deterministic, synchronous, 1-D States mapped to colors (CMYK) Software Rule generation (Mike Smith) Rule parser (OCCAM --> Netlogo) Rule visualizer - Netlogo Width of array = 300 cell torus Time steps per screen = 300 ticks (option to continue) I decided to use cellular automata for the image generation. While classifying C Ay's isn't exactly a science, as we established on Tuesday, C A's do provide a way to visualize different dynamic regimes and classify some of them unambiguously. C Ay's have also been used extensively by artists to generate interesting images. Additionally, C Ay's are an artificial life technique I hadn't explored previously. The C Ay's were K equals 4, N equals 5, deterministic, synchronous, and one dimensional, as described by Langton in the first paper we red for this class. The C ay states would be mapped to colors to create a colorful space time diagram. I picked the print primaries, cyan, magenta, yellow, and black for the states. In order to generate the images, I needed something which generated the C ay rules, and Mike did that part, saving me oodles of time. So then I just needed something that would parse those rules, which were in oc am format, into a format suitable for a visualizer. I decided to use NetLogo for the visualizer, which enabled me to generate large, complex C ay images really easily. The width of the world was a 300 cell tore us, and the height of the time steps was 300, with the ability to continue for another 300 ticks, indefinitely.

11 Image Generation: Visual Design Considerations Reducing confounders Mapping states to colors Different main color per image per regime Care not to misrepresent dynamics Consistent size, scale 550 x 550 px total image 2 x 2 px cells Increasing interestingness Optical mixing Selection for distinctiveness Photoshop filters Blur 2x Glowing Edges 1x There ended up being two points of view I had to take for every decision I made on this project. The first as a software engineer, and the second as a visual designer. I suppose this is fitting considering the topic of the project. As a primary visual design consideration, I needed to reduce confounders. The choice of mapping states to colors instead of to shapes or textures was to avoid the introduction of visual element confounders. I decided to use a different main color per image per regime, in an attempt to reduce color as a confounder. Care needed to be taken not misrepresent the dynamics by introducing compositional variation unrelated to regime. Images all had to have the same size, 550 by 550 pixels, and scale of 2 by 2 pixels each cell, to reduce size and scale as confounders. I also needed to increase the "interesting ness" of the images to avoid participant fatigue. The choice of primaries for cell states should cause "optical mixing," when separate colors near each other appear like the color that would be produced if those colors were mixed together. Optical mixing gives the appearance of more than four colors in the C Ay. It was also important that each image was as distinct as possible to prevent fatigue, which could confound preference, while retaining enough similarity for comparison of images to be meaningful. The color differences helped with this, as well as deliberate selection by me of images for their distinctness from each other. To make the images produced by NetLogo look less like old video game graphics, the blur and glowing edges filters were applied in Photoshop. The same filters were applied to all images, and were ones which remove the pixel ation without modifying the underlying structure and visible dynamic of the C Ay.

12 Image Generation: Measurement and Classification Not used Lambda poor measure Transient length insufficient time to implement Regime based on visual inspection Apparent - over 300 time steps Actual - over extended (eg 1200) time steps Complexity Ranking First by regime Second by apparent visual complexity Partially subjective Several measurements were considered for quantifying regime and complexity. Lambda was discarded quickly as having little relation to the dynamics. Transient length may have been a good metric, but there was insufficient time to program a transient counter. Visual inspection was the easiest way to classify a C ay dynamic regime. I considered both the dynamic displayed over the 300 ticks of one visualization, as well as the dynamic displayed over extended time, up to 1200 ticks. "Complexity Ranking" was my own partially subjective measure to rank images based both on regime and visual complexity, with regime as the priority criteria. While visual complexity is relatively subjective, this was the best way I could come up with quickly to get a more fine grained way to sort images. I felt it was important to do this as there was some wide variation between images used in each regime, like a square of a solid color is going to be responded to differently than multicolored stripes, even though both may classify as a fixed point attractor.

13 Image Generation: Implementation Since that was a lot of blah blah blah, here's the quick, illustrated version of how I implemented the image generation. I got rules from Mike's python program, in oc am format. I input the rules to a perl parser that output them in netlogo format. Those rules were input into the netlogo visualizer, and I visually assessed the dynamic, noting down the behavior shown. If the image was not useful for the survey, I went to the next rule. If the image was useful for the survey, I had Netlogo output it as a graphics file, which I then ran through the photoshop filters. Then I moved onto the next rule. I evaluated about 110 rules total, and used 14 of them to produce 15 images.

14 Image Generation: Raw Images These are the images as output by NetLogo, all but the second 33 with cyan as the primary color. In this picture, the images are arranged in rows by regime and in order of complexity ranking. The regime sorting is based on the dynamics shown in the image, not the dynamics of the rule over time. For example, rule 37 is actually a fixed point rule, even though its long transient makes it initially appear like edge of chaos. The edges between regimes blur a bit, for example at rule 33 which is a limit cycle rule will sometimes show chaotic behavior for a moment early on. Rule 62 remains somewhat ambiguous whether it is chaotic or edge of chaos, it's like edge of edge of chaos.

15 Image Generation: Final Images And here are the same rules, just produced in different primary colors and run through the photo shop filters. These are the images you saw if you ended up taking the survey.

16 Survey: Technical Considerations Internet based Hosted on personal site, created in HTML::Mason Data considerations (some regrettable) Data storage in flat CSV table One preference ranking per row Session information not saved No back button (would require session info.) Measurement Rank 1 6 (1 = strongly disliked, 6 = strongly liked) So. The survey. It was internet based, hosted on my personal site, and used the html mason web ay pee I. Survey data was stored in a flat C S V table on the web server, with image names as column names, and one preference ranking per row. The latter was in part because I did not save session information for individual users. No back button was provided, as it required session information, and I wasn't sure I wanted a back button any way, as it would destroy that "gut reaction" thing. The scale of measure for preference was a ordinal scale of 1 to 6, with 1 being strongly dislike and six being strongly like. I did not use a scale with a neutral point in order to force users to choose an actual preference.

17 Survey: Visual Design Considerations Reducing confounders Neutral background White space around image Images displayed one at a time Images displayed in a random order Increasing Usability Total 15 images Sufficient number data points 5 10 minutes to assess Image count down Instructions at top of every page 550 K or less image file size Radio buttons / simple interface Visual design considerations to reduce confounders included a neutral background for the survey pages and a swath of white space around each image, to avoid the surrounding page from influencing perception of the image. Images were displayed one at a time so as to avoid having them influence each other. Images were displayed in a random order to reduce sequence related confounders. Even though sequence might influence participants, for example they may all get bored and hate what they see last, at least each participant will be influenced on different images. Usability of the interface also had to be considered, especially to increase the chances of a participant completing the whole survey. The amount of images selected for the survey, 15, was intended to make for the shortest possible survey with a sufficient number of data points for each regime. Testers said it was taking between 5 and 10 minutes, usually on the low end, to assess the images in the survey, which seemed a reasonable amount of time to ask people to give up for uncompensated participation. An image count down helped users to know when they were almost at the end rather than giving up only a few images from the end. Instructions were repeated at the top of every page. Image files were kept to 550 K or less to minimize download time. Radio buttons and a really simple interface made it as hard as possible for a user to do something wrong.

18 Survey: Participant Recruitment Dissemination LiveJournal personal blog and communities Syscdiscuss elist A few others Let Internet dynamics do the rest! Time line 5/20 initial recruitment message 5/20 survey open 6/02 survey close 6/06 results posted I created a recruitment message for survey participants, which you probably all saw. I also posted the message to my personal blog, and to several lieve journal communities such as all complexity and math art. I also posted to the systems science discussion list, my research group, and some random other people. I then let the dynamics of the internet take over, hoping there would be sufficient viral activity to do my recruiting for me. The time line for the survey was only two weeks. I started posting recruitment materials on the 20th when the survey opened. I closed the survey on the 2nd. I'll be posting a copy of this presentation and summary of results for interested participants tomorrow.

19 Survey: Implementation Once more to give something besides a bunch of blah blah blah, here's the summary of how the survey worked. Participants would arrive at the instruction page, submit to the next page, and a random one of the 15 images that the participant has not yet seen would be selected, and displayed in the survey page. The participant would then select a preference, submit it, and the database would be updated. If it's not the last image, another random image that hasn't been seen is selected to display. If it is the last image, the participant is sent to a thank you page, reminding them results will be posted on the 6th.

20 Survey: Instruction Page Here's the instruction page, which you may have seen. Note that these example images were not used in the survey itself.

21 Survey: Interface Here's the survey interface. As a bit of foreshadowing, this particular screen shot shows the image that was ranked highest in preference order out of all of them.

22 Results: Population Description Dissemination tracking Additional personal blogs Additional community blogs Additional mailing lists Population Internet users May be more likely to have science, complex systems, or art backgrounds May be more likely to have education But may also not (e.g. Damn Portlanders) So who were these participants. I was able to track dissemination to additional personal blogs, a few additional community blogs including some large and diverse ones, and one additional mailing list. I lost track after that. Three cheers for internet viral dynamics. So the participants were all definitely internet users. Based on the places where I initially posted recruitment, participants may be more likely to have science, complex systems, or art backgrounds, and may be educated. However, the fact that the recruitment was posted on some very large, general boards such as damn portlanders means there was probably some real diversity to the participants population as well.

23 Results: Participation Statistics Number of participants 263 participated 208 lower bound for survey completion 257 upper bound for survey completion 230 approx. no. who completed the survey Number of preference points 775 Fixed Point (note: 3 images) 1040 Limit Cycle 1035 Edge of Chaos 1040 Chaos Considering how non aggressive I was about recruitment, the miniscule window of time the survey was open for, and the fact that participants were uncompensated, I was really impressed by how many individuals responded. 263 total participated, and at least 208 completed the survey and at most 257 completed the survey. So I had like 230 people actually complete the survey. The total amount of preference data points for each regime were about the same, remembering that there were only three fixed point images, and four of the others. Now onward to the real, actually exciting, results.

24 Results: Average Preferences by Regime Fixed Point = Limit Cycle = Edge of Chaos = Chaos = These are the average preferences for each of the regimes. While fixed point, limit cycle, and chaos hover around the average, edge of chaos is way higher than the others. The extremes of fixed point and chaotic were least liked.

25 Results: Histograms by Regime This diagram shows histograms of the preference distribution for each regime. The fixed point regime has mostly scores in the strongly disliked and disliked range, and very few strongly liked. The preferences for limit cycle and chaotic dynamics appear to follow a slightly skewed Bell distribution, although chaotic may be bi modal, probably reflecting the high preference for the edge of edge of chaos image 62, which remember I struggled over whether it was in the edge of chaos or chaotic regime. The preference distribution for the edge of chaos regime is clearly different, with a median at 5, and unlike the other regimes has a large amount of responses in the strongly liked category. Very few people strongly disliked, disliked, or even slightly disliked this regime. People definitely responded differently to the edge of chaos set of images.

26 Results: Average Preferences by Image This graph plots the average preference for each image sorted by Complexity Ranking, with the boundaries of each regime labeled on the chart. People seem to like the more complex images on the orderly end, and the less complex images on the chaotic end, and definitely prefer the images which exhibit both orderly and chaotic dynamics. This is consistent with my hypothesis. This plot shows the over all trend from dislike to like and back to dislike as the images move through each regime.

27 Results: Avg. Preferences by Image & Regime This graph plots each image by regime and sorted by Complexity Ranking, by average preference. This is essentially the same as the previous graph, but illustrates the point in a different way as a side-byside comparison for the images in each regime. Because the images are sorted by Complexity Ranking the most ordered images are at the left, and the most chaotic images are at the right. The most orderly and most chaotic images rank low, as do the more orderly limit cycle images. The most stereotypically "edge of chaos looking" image, remember the one that was illustrated in the survey screen shot, ranks at the top. Here's that image 62 which may be accounting for the bi modal looking distribution in the chaotic regime.

28 Conclusions Found evidence to support hypothesis that people will find images which exhibit Edge of Chaos dynamics more aesthetically appealing than images with other types of dynamics. Additionally found evidence that the complexity of the image, in combination with its regime, may influence aesthetic appeal in favor of Edge of Chaos. This project was fantastically fun! The results of the survey statistics I just showed support the hypothesis that people will find images which exhibit edge of chaos dynamics more aesthetically appealing than images with other types of dynamics. Additionally, it appears that the complexity of an image, in combination with its regime, may influence the aesthetic appeal in favor of an edge of chaos sweet spot, neither too complex nor too simple. Maybe most importantly, this project was super fun and I'm extremely pleased with it.

29 Future Work More analysis of current data More rigorous complexity measure Alternative complexity / chaosity measures Tests for statistical significance Dealing with outliers Larger scale investigation Save session info and keep responses together More sophisticated statistics (e.g. cluster analysis) Participant information? More exhaustive review of current literature Which doesn't mean the same thing as this project being perfect. It was sort of a rough draft, and there is more I can do both with the current information I have and with extending the work. From the time line I gave, you may have noticed I only had one day in which to perform any sort of analysis on the data, oops. Some things to be done right now include applying a more rigorous measure of complexity to the images, of which there are many that could work, and sorting the images by alternate complexity and chaosity measures. Tests for statistical significance still need to be done. Also, visual inspection of the data shows some obvious outlier behaviors in participant responses, such as people who answered strongly like on every image. Maybe this sort of thing should be considered more deeply. Beyond the current data set, the most regrettable part of this project for me was not keeping participant responses together. This means I'm unable to do any real sophisticated statistical analysis, such as cluster analysis, that could provide extremely interesting information about the survey data. I had to make the survey super anonymous as there was no time for working with the I R B, but if the study was redone more carefully in the future, it might be interesting to collect some additional information from participants, such as whether they were assessing artistic or mathematical aesthetics in their definition of appeal. Finally, there is a massively extensive literature on the intersection of art and artificial life. More time spent in that literature could stimulate ideas both for extending this work, and for applying artificial life to my own art.

30 Game of Life Dora Raymaker ~ draymake@pdx.edu Which I will end with. This is a painting called game of life, and it's actually got some gliders in it. The intersection of art and artificial life is extremely (pun intended) fertile. Perhaps this is because, I feel anyway, art actually is an artificial life methodology, as art by its very definition is a way of modeling "life as it is" as "life as it could be." I will take questions now. I have references on the next slide as well, if you'd prefer to look at that. Let me know if you want to see them.

31 References Langton, C. G., "Life at the Edge of Chaos", ALife II, 41-91, Casti, J., A. Karlqvist. Art and Complexity. Elsevir, Amsterdam Whitelaw, M. Metacreation: Art and Artificial Life. The MIT Press, Cambridge DiPaola, S., L. Gabora. "Incorporating Characteristics of Human Creativity into an Evolutionary Art Algorithm." GECCO'07, July 7-11, 2007, London, England, United Kingdom. Dorin, A. "Art and Artifice." Artificial Life, v. 9, n. 1, p Langton, C. G., Artificial Life, Artificial Life (Vol. VI, Santa Fe Institute Studies in the Sciences of Complexity), Langton, C., ed., Addison- Wesley, Redwood City CA, 1-47, Sims, Karl, Genetic Images Karl Sims. June 5, <