Evolution of learning is key to better artificial intelligence

Evolution of learning is key to better artificial intelligence
Anselmo Pontes, MSU computer science researcher. Credit: Michigan State University

Since "2001: A Space Odyssey," people have wondered: could machines like HAL 9000 eventually exist that can process information with human-like intelligence?

Researchers at Michigan State University say that true, human-level intelligence remains a long way off, but their new paper published in The American Naturalist explores how computers could begin to evolve learning in the same way as did—with implications for many fields, including artificial intelligence.

"We know that all are capable of some form of learning, we just weren't sure how those abilities first evolved. Now we can watch these major evolutionary events unfold before us in a virtual world," said Anselmo Pontes, MSU computer science researcher and lead author. "Understanding how learning behavior evolved helps us figure out how it works and provides insights to other fields such as neuroscience, education, psychology, animal behavior, and even AI. It also supplies clues to how our brains work and could even lead to robots that learn from experiences as effectively as humans do."

According to Fred Dyer, MSU integrative biology professor and co-author, these findings have the potential for huge implications.

"We're untangling the story of how our own cognition came to be and how that can shape the future," Dyer said. "Understanding our own origins can lead us to developing robots that can watch and learn rather than being programmed for each individual task."

The results are the first demonstration that shows the evolution of associative learning in an artificial organism without a brain.

"Our inspiration was the way animals learn landmarks and use them to navigate their environments," Pontes said. "For example, in laboratory experiments, honeybees learn to associate certain colors or shapes with directions and navigate complex mazes."

Since the evolution of learning cannot be observed through fossils—and would take more than a lifetime to watch in nature—the MSU interdisciplinary team composed of biologists and computer scientists used a digital evolution program that allowed them to observe tens of thousands of generations of evolution in just a few hours, a feat unachievable with living systems.

In this case, organisms evolved to learn and use to help them navigate the environment and find food.

"Learning is crucial to most behaviors, but we couldn't directly observe how learning got started in the first place from our purely instinctual ancestors," Dyer said. "We built in various selection pressures that we thought might play a role and watched what happened in the computer."

While the environment was simulated, the evolution was real. The programs that controlled the digital organism were subject to genetic variation from mutation, inheritance and competitive selection. Organisms were tasked to follow a trail alongside signals that—if interpreted correctly—pointed where the path went next.

In the beginning of the simulation, organisms were "blank slates," incapable of sensing, moving or learning. Every time an organism reproduced, its descendants could suffer mutations that changed their behavior. Most mutations were lethal. Some did nothing. But the rare traits that allowed an organism to better follow the trail resulted in the organism collecting more resources, reproducing more often and, thus, gaining share in the population.

Over the generations, organisms evolved more and more complex behaviors. First came simple movements allowing them to stumble into food. Next was the ability to sense and distinguish different types of signals, followed by the reflexive ability to correct errors, such as trying an incorrect path, backing up and trying another.

A few organisms evolved the ability to learn by association. If one of these organisms made a wrong turn it would correct the error, but it would also learn from that mistake and associate the specific signal it saw with the direction it now knew it should have gone. From then on, it would navigate the entire trail without any further mistakes. Some organisms could even relearn when tricked by switching signals mid-trail.

"Evolution in nature might take too long to study," Pontes said, "but evolution is just an algorithm, so it can be replicated in a computer. We were not just able to see how certain environments fostered the evolution of learning, but we saw populations evolve through the same behavioral phases that previous scientists speculated should happen but didn't have the technology to see."

Other MSU co-authors include Robert Mobley, Charles Ofria and Christoph Adami. This project was developed through the BEACON Center for the Study of Evolution in Action, which brings together biologists, computer scientists and engineers to illuminate and harness the power of .

"Pontes and colleagues have evolved associated learning in a computer from the raw ingredients of mutation, inheritance and competitive selection," said George Gilchrist, program director at the National Science Foundation, which funds the BEACON science and technology center. "This opens the door to creating artificial intelligence systems without the limitations imposed by human design."


Explore further

Online tool speeds up evolution education

More information: Anselmo Pontes et al. The Evolutionary Origin of Associative Learning, The American Naturalist (2019). DOI: 10.1086/706252
Journal information: American Naturalist

Citation: Evolution of learning is key to better artificial intelligence (2019, September 20) retrieved 22 October 2019 from https://phys.org/news/2019-09-evolution-key-artificial-intelligence.html
This document is subject to copyright. Apart from any fair dealing for the purpose of private study or research, no part may be reproduced without the written permission. The content is provided for information purposes only.
356 shares

Feedback to editors

User comments

Sep 21, 2019
"Researchers at Michigan State University say that true, human-level intelligence remains a long way off..."

Sure, 3-4 years max. More like 2 years. The AI that matters would be deeply classified operations, funded to the tune of 100's of billions of dollars in the U.S., China, Japan, Russia.
But the notion of " 'true' human-level" intelligence is misinformative at best, disingenuous at worst. Intelligence proper, the capacity to map the physical, temporal, and abstract elements of the world into manipulatable tokens inside a sort of scale-model 'world', is not constrained by cerebral architecture.
As a matter of fact, no programme for AI funded by state actors could find itself answerable to bureaucrats and legislators and paymasters under such a delimiting...artificial barrier.

Want to hog-tie the opposition and pin them down? The best route is to   s t e a l   ideas for effective AI realization, whether from humans, portia spiders, or, whatever already works for machines.

Sep 22, 2019
but evolution is just an algorithm, so it can be replicated in a computer


Evolution is a stochastic process, not an algorithm, and by definition cannot be replicated. You can't make "random" happen twice, nor can you produce randomness with an algorithm.

For an algorithmic pseudo-random system, the process is deterministic and you may or may not evolve certain features simply because of the seed you give the random number generator.

Sep 28, 2019
Besides, like the other article pointed out, the way computers deal with numbers (floating point representation), the results you get are not scale-invariant.

In other words, features that appear at one scale of numeric representations cannot happen at a much larger or smaller scale because the numbers used to represent these features do not behave the same way. This breaks down a lot of features of our reality, such as self-similarity across different scales, which means the computer simulation behaves in a fundamentally different way - it's like a different universe with different laws of physics entirely.

Evolution takes advantage of ALL the features of the environment that have any effect on the outcomes, which means this different universe inside a computer may lack some of the features that make "intelligence" work in the real world. In fact, it most certainly does - see the Chinese Room Argument.

Sep 28, 2019
That's not to say computer programs cannot evolve to behave in ways that look intelligent, but that computer programs are still programs. One cannot describe a list of instructions as being "intelligent" without resorting to some magical leaps of imagination.

Sep 28, 2019
Another way of looking at the point is observing that "intelligence" requires the presence of choice. If your actions are forced, you are not making choices and therefore talking about intelligence is meaningless.

A program that is using a random number generator to vary its response is still completely defined in advance. In theory, you could determine the responses of the program to any sequence of events if you know the RNG algorithm and the seed. The program really has no choice - which means it cannot be intelligent. Using a real source of randomness, such as seeding the RNG with observations of a radioactively decaying particle, does not bypass the problem because the system hasn't changed. You simply delegate the choice.

That's why there's limited value in studying computer models of intelligent organisms - because they aren't intelligent, and they aren't organisms. They don't "learn" - the program can only evolve in one way each time.

Please sign in to add a comment. Registration is free, and takes less than a minute. Read more