The research, actually over 5 years old is about evolutionary banner design. That is, using genetic algorithms to design advertisement banners in order to maximize their click-through rates.

The research was conducted on the e-commerce site boxman.se during an 11 day period in 1998. The basic layout of a simple banner design was broken into 4 areas and each area could be used to display various different elements about the product, such as the artist’s name, the title of the album, cover art, artist images and slogans. 5 different font types and 2 colors were also among the things that the genetic code could play with. (See more about genetic algorithm’s in yesterday’s entry.)

Using these features, an original set of 20 banners was randomly generated. The fitness of a banner was then measured by its click-through rate and the “fittest” designs used to breed the next generation, etc. For comparison, part of the banners displayed in the research were not a part of the evolutionary process, and hence remained the same during the entire period (for details refer to this paper).

The results were quite interesting. While the non-evolving banners had from day 1 a steady click-through rate of about 0.7%, the evolving banners grew over 9 generations from 0.95% to 1.41% (average for the entire generation). Same numbers, looking only at the numbers for the 5 most successful banners of each generation rose from 1.3% to 2.6% – doubling the click-throughs and more than tripling the rate for that stationary designs. Not bad!

Interestingly enough the last generation showed several trends, most notably that including the phrase “click here” plays a big role in the success of a banner. This is in line with earlier results from other academic research (oh, the simple creatures we are).

Given these results, I’m surprised that this practice doesn’t seem to be used for today’s advertisement banners, despite the obvious value. I also wonder if similar methodology could be used e.g. in usability design or even graphical design and the layout of web pages. The beauty of it being that users’ typical behavior is being used as a measurement of fitness, there is no additional survey or “Did you find this helpful?” links to click.

Details:
For a quick overview, you can browse through Gatarski’s presentation slides on the project from COTIM-99.

Details can be found in Gratarski’s papers from Generative Art ’98 and COTIM-99. The GA paper has more details on the design elements, while the COTIM one spends more time on methodology and results.

Using genetic methods, people try to “breed” software that best meets the task that is to be solved (the goal), just like nature’s evolution processes breed animals that are best fit to survive in their natural environment. In this process nothing is “forbidden”, the individuals can try whatever means available to them to seek the optimal solution, and in the case of a computer environment that can mean exploiting unknown or at least not-intended-to-use flaws in the environment’s design.

This spring I wrote an essay called “A mind emerges” in a Philosophy class. In the essay I mentioned this phenomenon and used two examples:

1. Hardware chip flaw: In an experiment to train a neural network to distinguish between different frequencies of sound, Professor Adrian Thompson of the University of Sussex ran into an interesting problem. When he had trained the network and was satisfied with its accuracy rate (which was almost 100%) he took the network that had evolved and started to deduct the parts of the network that did not contribute to the task of distinguishing the frequencies. Using various methods, he removed parts of the network and simplified it.

   But at one point something unexpected happened. Removing a link that did not seem to contribute anything to the real functional circuitry made the program fail. The link wasn’t even a part of the circuit between the network’s input and output! Obviously, this did not make any sense at first, but after doing some further experiments he found the reason. The computer chip he was running the neural network on was flawed. When the network that he had successfully trained on this particular chip was copied on to another chip that was supposed to be identical to the previous one, the network didn’t work at all. When training the system, the network had found a way to use the flaw in the chip to its advantage.

2. Software physics model flaw: Another similar story comes from a work by GenArts genius Karl Sims on simulated creatures that were evolved in a simulated environment. Sims used genetic methods to develop means of locomotion for his computer creatures (not unlike the snake). But a few of Sims’ creatures developed a very strange and what whould under normal circumstances have been impossible means to move. This turned out to be a result of a flaw in the implementation of the physics in Sims’ simulated environment. The evolving creatures had spotted the flaw and exploited it to their advantage. (Discussion on this exploit can be found in ‘Frontiers of Complexity’ by Peter Coveney and Richard Highfield, see: Spreading the Cause).

Both of the above samples are very interesting and are good examples of this exploiting behavior. Now what I’m wondering about is what this tells us about methods and solutions that have evolved in the “real” nature? The fact that animal organs sometimes evolve to serve a different purpose than before seems to be of similar nature. Exploitation of this sort certainly leads to more complex organisms rather than simpler, as a minor “defect” in an individual could be exploited to serve as an organ with a specific purpose. And what might this mean for as complex systems as the human brain? Without doubt helpful defects on the evolutionary road have been used or as I put it in my aforementioned essay:

“[Exploiting of this sort] shows that a neural network cannot be abstracted from the environment in which it functions. We cannot necessarily simulate the functionality of the brain just by simulating the neuron functionality. Answers might be found in the neural network’s environment, e.g. the glia cells (that are mainly believed to be a support system for the neural structure) or even in how the skull affects the brainwaves or how the blood flowing through the brain carries different chemicals. This might mean that even though the brain’s neural network carries out the main functionality of the brain, nature might have found ways to exploit “defects” in its own design for its benefit in a similar way that Thompson’s neural net exploited the flawed chip or Sims’ creatures exploited the flaw in the physical simulation.”

Links:
Karl Sims’ Evolved Virtual Creatures
Video of Sims’ creatures (don’t miss this one, it is fantastic)