Charles Darwin, in his theory of evolution, argued that all living species have evolved over time from a single or few common ancestors. Thus, animals have an embodied intelligence: the adaptation of their morphology to their environment has allowed them to accomplish complex tasks. Researchers at Standford University wondered whether an AI could evolve in the same way as a living being, and their research allowed them to create one that transforms according to the complexity of their environment. The results of their study titled: “Embodied intelligence via learning and evolution” were published in the journal Nature Communications.
The aim of this study is “ to elucidate some principles governing the relationships between environmental complexity, evolved morphology and the learnability of intelligent control”. Fei-Fei Li, a member of the research team and co-director of the Standford Institute for Human-Centered AI (HAI) states:
“We are often so focused on the fact that intelligence is a function of the human brain and neurons in particular. Viewing intelligence as something that is physically embodied is a different paradigm. “
An in silico playground
For the study, the researchers created a computer-simulated playground where arthropod-like agents called “unimals” (for universal animals) will learn and be subject to mutation and natural selection. then studied how having virtual bodies affected the evolution of the unimals’ intelligence.
To successfully scale simultaneously the creation of embodied agents on 3 axes of complexity: environmental, morphological and control, the researchers designed a deep evolutionary reinforcement learning (DERL) algorithm.
To understand the evolution of embodied intelligence, the team varied not only the body shapes of the unimals, but also their training environments and the tasks they performed. Study co-author Surya Ganguli, associate professor of applied physics in the School of Humanities and Sciences and associate director at HAI, said:
“And all of these variables were much more complex than in previous work, so it allowed us to look at many more scientific questions than before. “
To maintain the diversity of unimals and reduce the computational cost of these simulations, the researchers opted for a tournament-style Darwinian evolutionary scheme that allowed them to ensure that each unimal morphology had a chance to succeed and be passed on to the next generation.
Each simulation started with 576 unique unimals and the same neural architectures and learning algorithms. During training, each moved either on flat terrain or on more challenging terrain with stair steps, smooth hills, or blocky ridges. They then entered a tournament against three other unimals trained under the same conditions as they were.
The winner was chosen to produce a single offspring that underwent a single mutation involving limb or joint modifications before facing the same tasks as its parents. All unimals (including the winners) competed in multiple tournaments, aging only as new offspring emerged. The researchers stopped the simulation when they had 4,000 different morphs. The surviving unimals had then gone through, on average, 10 generations of evolution, and the successful morphs were surprisingly diverse, including bipeds, tripeds, and quadrupeds with and without arms.
Results of the study
The researchers selected the 10 best performing unimals from each environment and trained them on eight new tasks: navigating around obstacles, manipulating a ball, pushing. a box up a slope… They found that the unimals who moved on variable terrain performed better than those who moved on flat terrain but were outperformed by those who manipulated a box on variable terrain. On the other hand, at this stage, they were learning twice as fast as their first ancestor.
These results could be of interest to the robotics industry for the creation of multi-tasking robots.
Article source: https://doi.org/10.1038/s41467-021-25874-z.
Translated from L’Intelligence Artificielle peut elle évoluer à l’image des êtres vivants ?
