Sign In

Communications of the ACM

ACM TechNews

Researchers Are ­sing Darwin's Theories to Evolve Ai, So Only the Strongest Algorithms Survive

View as: Print Mobile App Share:
Biology, with some artificial help.

Modern artificial intelligence is built to mimic nature. The fields main pursuit is replicating in a computer the same decision-making prowess that humankind creates biologically.

Credit: Siphiwe Sibeko/Reuters

Researchers at Google Brain and OpenAI are applying Darwinian principles of evolution to advance artificial intelligence (AI).

Google's neuroevolution project trained 1,000 image-recognition algorithms on deep-neural networks to recognize specific images. The more accurate algorithms were then copied and "mutated" to see if their clones' accuracy would improve, with such mutations allowed to survive and eventually achieve 94.6% recognition accuracy.

Meanwhile, OpenAI's research focused on using "worker" algorithms to train a master AI to perform an unknown task. The evolutionary AI tracks how workers learn, thus learning how to extract more insight from the same amount of data. The workers played Atari and reported their scores to the master. The highest-scoring algorithms were copied and randomly mutated, then put back into rotation so subsequent mutations could be copied or deleted, depending on their scoring prowess.

OpenAI's approach is considered to be closer to evolution's true biological function.

From Quartz
View Full Article


Abstracts Copyright © 2017 Information Inc., Bethesda, Maryland, USA


No entries found

Sign In for Full Access
» Forgot Password? » Create an ACM Web Account