A computer science team at The University of Texas at Austin has discovered that virtual mass extinctions push robots to evolve more quickly and efficiently, suggesting that mass extinctions speed up evolution by encouraging new features and abilities in surviving lineages.
In the world of the living, mass extinctions are associated with utter destruction and the loss of a significant amount of genetic material. This is something that’s generally seen as a negative, especially since surviving populations may be small, leading to inbreeding and eventually more extinctions.
However, some evolutionary biologists hypothesized that such events are actually more positive, as generally only the most evolvable species survive—meaning evolution can leap forward as they fill in the gaps left behind by now defunct species.
Or, as co-author and computer scientist Risto Miikkulainen phrased it, “Focused destruction can lead to surprising outcomes. Sometimes you have to develop something that seems objectively worse in order to develop the tools you need to get better.”
In order to test this theory, the researchers used simulated robot brains known as artificial neural networks (ANNs) on which scientists can use evolution-inspired algorithms in order to help the “brains” improve at a task from one generation to the next.
Evolving computer-simulated robot legs
In a computer simulation, they connected several ANNs (“brains”) to simulated robotic legs with the goal being that evolution would allow the legs to walk smoothly and stably. Like with real evolution, random mutations were introduced in the ANNs, and many niches were created for the ANNs to fill as they evolved. (Similar to biology, a niche here referred to a behavioral specialization of a robot.)
The ANNs were tested in six conditions: Control, where no extinction occurred; Extinctions 300, 600, and 900, where 90% of the population was randomly killed off every 300, 600, or 900 generations, respectively; and Extinction Random, where mass extinctions killed off 90% at a random interval between 300 and 900 generations. Each condition was tested for 5,000 generations.
The end result: the conditions including extinctions evolved the best solutions to the task of walking as compared to the control. Further, in the extinction simulations, each extinction event resulted in an indirect selection for more evolvable individuals—thereby selecting for the lineages with the most potential to produce new behaviors. Of all the conditions, Extinction 300 resulted in lineages with the greatest evolvability.
While this is extremely interesting in the context of evolution, it has practical applications as well. For example, this could lead to the development of robots that can better search for survivors in earthquake rubble, or even navigate minefields.
(Image: At the start of the simulation, a biped robot controlled by a computationally evolved brain stands upright on a 16 meter by 16 meter surface. The simulation proceeds until the robot falls or until 15 seconds have elapsed. Credit: Joel Lehman)