Computer scientists find mass extinctionscan accelerate evolution12 August 2015

At the start of the simulation, a biped robot controlled bya computationally evolved brain stands upright on a 16meter by 16 meter surface. The simulation proceedsuntil the robot falls or until 15 seconds have elapsed.Credit: Joel Lehman

A computer science team at The University ofTexas at Austin has found that robots evolve morequickly and efficiently after a virtual mass extinctionmodeled after real-life disasters such as the onethat killed off the dinosaurs. Beyond its implicationsfor artificial intelligence, the research supports theidea that mass extinctions actually speed upevolution by unleashing new creativity inadaptations.

Computer scientists Risto Miikkulainen and JoelLehman co-authored the study published today inthe journal PLOS One, which describes howsimulations of mass extinctions promote novelfeatures and abilities in surviving lineages.

"Focused destruction can lead to surprisingoutcomes," said Miikkulainen, a professor ofcomputer science at UT Austin. "Sometimes youhave to develop something that seems objectivelyworse in order to develop the tools you need to getbetter."

In biology, mass extinctions are known for beinghighly destructive, erasing a lot of genetic materialfrom the tree of life. But some evolutionarybiologists hypothesize that extinction eventsactually accelerate evolution by promoting thoselineages that are the most evolvable, meaning onesthat can quickly create useful new features andabilities.

Miikkulainen and Lehman found that, at least withrobots, this is the case. For years, computerscientists have used computer algorithms inspiredby evolution to train simulated robot brains, calledneural networks, to improve at a task from onegeneration to the next. The UT Austin team'sinnovation in the latest research was in examininghow mass destruction could aid in computationalevolution.

In computer simulations, they connected neuralnetworks to simulated robotic legs with the goal ofevolving a robot that could walk smoothly andstably. As with real evolution, random mutationswere introduced through the computationalevolution process. The scientists created manydifferent niches so that a wide range of novelfeatures and abilities would come about.

After hundreds of generations, a wide range ofrobotic behaviors had evolved to fill these niches,many of which were not directly useful for walking.Then the researchers randomly killed off the robotsin 90 percent of the niches, mimicking a massextinction.

After several such cycles of evolution andextinction, they discovered that the lineages thatsurvived were the most evolvable and, therefore,had the greatest potential to produce newbehaviors. Not only that, but overall, bettersolutions to the task of walking were evolved insimulations with mass extinctions, compared withsimulations without them.

1 / 2

Practical applications of the research could includethe development of robots that can better overcomeobstacles (such as robots searching for survivors inearthquake rubble, exploring Mars or navigating aminefield) and human-like game agents.

"This is a good example of how evolution producesgreat things in indirect, meandering ways," explainsLehman, a former postdoctoral researcher inMiikkulainen's lab, now at the IT University ofCopenhagen. He and a former student ofMiikkulainen's at UT Austin, Kenneth Stanley,recently published a popular science book aboutevolutionary meandering, "The Myth of theObjective: Why Greatness Cannot Be Planned.""Even destruction can be leveraged for evolutionarycreativity."

Provided by University of Texas at AustinAPA citation: Computer scientists find mass extinctions can accelerate evolution (2015, August 12)retrieved 13 August 2015 from http://phys.org/news/2015-08-scientists-mass-extinctions-evolution.html

This document is subject to copyright. Apart from any fair dealing for the purpose of private study or research, nopart may be reproduced without the written permission. The content is provided for information purposes only.

Powered by TCPDF (www.tcpdf.org)

2 / 2