This week’s BEACON Researchers at Work post is by University of Texas graduate student Dan Lessin and Nicole Lessin.
As an undergraduate in the 1990s, I was studying studio art, animation, and computer graphics at Harvard when I first came across a video that accompanied what we now know to be a landmark publication by Karl Sims. This work demonstrated evolved virtual creatures made of simple cube-like shapes performing lifelike behaviors in a physically simulated environment. Sims had simultaneously evolved the bodies and brains of these creatures to perform locomotion on land and through water, as well as a number of other behaviors. Some of them hopped, some pulled themselves along the ground, others paddled or snaked through the water, and the most compelling examples were even able to follow and navigate to a user-controlled light source. It was like spying on alien life forms inhabiting a new world. And they all looked different from one another, as each had its own body plan and style of motion which had emerged naturally from Sims’ process of simulated evolution.
Like others who have seen Sims’ work, I found these creatures extremely compelling. I would later learn that my reaction was an example of an effect called “perceptual animacy,” the phenomenon through which viewers attribute internal motivations and desires to even the simplest of geometric shapes if they display the right kind of movement. Maybe bringing even more behavioral complexity to these virtual creatures could elicit an even stronger response of this type.
When I returned to academia after many years to begin work on a PhD in Computer Science at the University of Texas at Austin, I was pleasantly surprised to find that, despite many extensions to Sims’ work in a number of different fields, nobody had yet demonstrated more complex behaviors than the light following that Sims had produced so many years earlier. So for my own research, I began to think about how to evolve creatures with more complex behaviors toward the end of generating even more compelling virtual creature content, such as one might find in a nature documentary or an internet cat video. For example, instead of just drifting smoothly to a target as in Sims’ work, new creatures might first move to, then strike at a target, then perhaps run away if scared, or instead maybe patiently stalk a target then pounce on it, as a real animal might do.
In contrast to Sims’ creatures, who learned their light following as one integrated skill, I set about adding behavioral complexity to my own through the use of a novel syllabus method, which was based on my time as an aviation instructor and the building-block method of instruction we employed to teach some of the more complex aspects of flying. For example, learning to hover in a helicopter is a complicated maneuver that requires the knowledge and integration of a number of different skills–using the collective to move up and down, the cyclic to move laterally, and the pedals to control heading–each of which is a challenge to master individually, let alone in combination with the others. So students are only allowed to have control over one of these elements at a time, until each is mastered on its own. Then they can learn to combine them with relative ease, resulting in a steady learning progression to achieve what might otherwise be an impossibly complex task.
Applying this kind of approach to evolved virtual creatures, I train them through a sequence of relatively simple subtasks in which the earlier skills make the later ones easier to achieve, and each task is only about as difficult as what would typically be required in a normal evolved virtual creature system. Using this method, I have so far been able to match, then approximately double the upper limit of behavioral complexity for evolved virtual creatures that had persisted for almost two decades since Karl Sims first set it in his 1994 publication. These results were presented at this year’s GECCO conference and are summarized in a 5-minute video available online [4]. The process is analogous to using a ladder rather than trying to improve your ability to jump. Through practice and the use of new techniques, you might eventually be able to jump twice as high as you can today. But if you can use a ladder, you can reach arbitrary heights with only a limited amount of effort at each step.
For more information about Dan’s work, you can contact him at dlessin at cs dot utexas dot edu.
