Karl Sims evolved simple blocky creatures to walk and swim (video). In the paper, he writes “For land environments, it can be necessary to prevent creatures from generating high velocities by simply falling over”—ISTR the story is that in the first version of the software, the winning creatures were those that grew very tall and simply fell over towards the target.
The designer’s intentions, implicit in the fitness criteria he specifies and the values he assigns to these criteria, become explicit when he intervenes to encourage “interesting” evolutions and prohibit “inelegant” ones (“3-D Morphology”, pp. 31, 29). For example, in some runs creatures evolved who achieved locomotion by exploiting a bug in the way conservation of momentum was defined in the world’s artifactual physics: they developed appendages like paddles and moved by hitting themselves with their own paddles. “It is important that the physical simulation be reasonably accurate when optimizing for creatures that can move within it,” Sims writes. “Any bugs that allow energy leaks from non-conservation, or even round-off errors, will inevitably be discovered and exploited by the evolving creatures,” (“Evolving Virtual Creatures,” p. 18). In the competitions, other creatures evolved to exceptionally tall statures and controlled the cube by simply falling over on it before their opponents could reach it (“3-D Morphology,” p. 29.) To compensate, Sims used a formula that took into account the creature’s height when determining its starting point in the competition; the taller the creature, the further back it had to start. Such adjustments clearly show that the meaning of the simulation emerges from a dynamic interaction between the creator, the virtual world (and the real world on which its physics is modeled), the creatures, the computer running the programs, and in the case of visualizations, the viewer watching the creatures cavort. In much the same way that the recursive loops between program modules allow a creature’s morphology and brain to co-evolve together, so recursive loops between these different components allow the designer’s intent, the creatures, the virtual world, and the visualizations to co-evolve together into a narrative that viewers find humanly meaningful...compared to artificial intelligence, artificial life simulations typically front-load less intelligence in the creatures and build more intelligence into the dynamic process of co-adapting to well-defined environmental constraints. When the environment fails to provide the appropriate constraints to stimulate development, the creator steps in, using his human intelligence to supply additional adaptive constraints, for example when Sims put a limit on how tall the creatures can get.
Karl Sims evolved simple blocky creatures to walk and swim (video). In the paper, he writes “For land environments, it can be necessary to prevent creatures from generating high velocities by simply falling over”—ISTR the story is that in the first version of the software, the winning creatures were those that grew very tall and simply fell over towards the target.
[Edited]
Karl ‘Sims’? Really?
Yes: “The Power of Simulation: What Virtual Creatures Can Teach Us”, Katherine Hayles 1999: