Evolutionary Algorithms and Open-Ended Evolution with Lisa Soros

Article by Hengtee Lim | December 02, 2020

One lesser known but growing area of AI research is the field of open-ended evolution and evolutionary algorithms. This area of research looks to model natural evolutionary systems in a computational environment in an attempt to create novelty. In essence, it’s the search to understand the important parts of the evolutionary process so computer algorithms can generate similar unique discoveries.

To learn more about how it works and the current research landscape, I spoke with Lisa Soros, a postdoctoral researcher at CrossLabs. Soros’ work focuses on evolutionary algorithms designed to look for novelty and discover new things, often in game environments. In this interview, we talked about how evolutionary AI works, current roadblocks to open-ended evolution, and why games are such a fertile environment for testing artificial intelligence.

 

What is Open-Ended Evolution?

The inspiration for open-ended evolution is the earth itself, and how it evolved over time to host a whole manner of complex systems, from plant life to animals, and of course, humans.

“With open-ended evolution, what we’re trying to replicate is a phenomenon like the big bang or the emergence of the first living cell,” says Soros. “On Earth this started from nothing, but also allowed for a host of complex things to exist—such as animals and plants—and a crazy amount of diversity within those living things.”

“So the idea is if we can figure out how that evolutionary process works and figure out what the most important parts are, then maybe we can make a computer algorithm that will generate a lot of interesting and complicated stuff that we can use for something useful. We’re not necessarily trying to create Earth again, that’s basically impossible, but we’re trying to figure out what are the most important parts to create a similar effect. “

Open-ended evolution is driven by questions that aim to understand the emergence of intelligence. As artificial life researcher Olaf Witkowski put it, “Think of the Earth as a box; how come when you shake it, and wait long enough, suddenly you get intelligent humans? Technology?”

In order to explore these questions, evolutionary algorithms are designed to look for things that are new and different, where new discoveries lead to more new discoveries, and on and on. This requires a system to be flexible, adaptive, and autonomous, but it also requires careful thought at the design level.

“If you’re looking at the real world and figuring out which important parts of a system can be turned into an algorithm,” says Soros, “then evolutionary algorithms take things like reproduction selection and mutation and apply them to anything you can represent in code as a data structure.”

 

Evolutionary Algorithm Design and Artificial Life

So how do you implement a system as large and far reaching in scope as evolution and biology? According to Soros, the size and scope of evolution itself is perhaps the biggest challenge when it comes to designing evolutionary algorithms.

“I think the main problem is that we’re trying to engineer something at a very large scale. We’re essentially trying to understand how evolution and biology happened by trying to reimplement it. But we don’t have 3.8 billion years to wait and see if we’ll get the same thing. On top of that, in terms of processing power a computer is not the same as the natural processing that happens on Earth. So when we create our abstractions, we’re implementing something that operates on a much smaller scale than natural evolution. The question then is how can we tell whether we’ve achieved the same dynamics that we’re copying from the real world?”

“We want our evolutionary system to find a lot of diverse things, whether it’s behavioral strategies or the generation of creative artifacts. And we want them to become more complex over time, so the question is, how do we know we’re getting the same kind of complexity? How do we know we’re getting the same diversity?”

The other challenge lies in learning to think outside of what we already know and see. We’re building systems that mimic behavior we see in nature and the evolution of the planet we live on, but that may not be the only answer available. “As humans, we’re trying to create something that is visually recognizable as being as interesting as Earth,” says Soros. “So another interesting challenge is in how to implement systems that don’t just directly copy that.”

 

The Relationship between AI and Games

Soros’ research to date has often centered on implementing systems around game structures. She says her love of games is a big motivator, and she hopes her research will one day support game developers. And it’s true to say that there’s a clear relationship between games and AI research. The two areas often seem intertwined. We’ve seen this in AI chess player Deep Blue, and more recently AlphaGo, which beat the world’s best Go player. But games are also used as a testing ground for AI systems, as we’ve seen in the use of the Arcade Learning Environment. I asked what Soros thought of the relationship.

“It’s interesting because there’s AI for games, and there’s games for AI. We can use games as an environment for testing AI [because] in general, playing games requires many skills we need for the real world; we need to react depending on the type of game, and we need to be able to think ahead to predict what will happen in the future. There’s also navigation and the task of simply looking at a screen and making sense of the colors and pixels, which is hard for computers. Compared to other domains for testing AI, games can also add an interesting complexity if there’s another player. [In this case] you’ve got to think about what the other player is going to do, how predictable they are, and be able to counter any adversarial things the other player might be doing.”

When it comes to games for AI, on the other hand, designing game structures is a simpler and more cost effective way to test a developing system.

“Games give us a computationally cheap way of testing things,” says Soros. “For example, it’s much easier to run a game than it is to build a robot, put it in the real world, and watch it do things. So [with games] we can prototype our ideas really fast and test them in a system that encapsulates some of the interesting parts of doing things intelligently in the world.”

 

Real World Applications for Evolutionary Algorithms and AI in Gaming

In her research, Soros and her fellow researchers have used evolutionary algorithms to generate playable game levels, and built a Sims simulator that discovers furniture layouts for a life simulation game. This is part of the way evolutionary AI research uses games with a variety of ways to win, as opposed to games with very clear singular goals, e.g. reach the end of a level. Through this research, Soros sees an opportunity for AI in the area of game testing and balancing.

“I’m hoping to continue working [in the future] on the game support tools that I’ve been researching. As somebody who plays games, I’d like to help improve the things I do for fun with my algorithms. [For example] developing algorithms that will help do automated testing for game designers who have a vision of the game they want to make. Such an algorithm would help protect the design to ensure you can’t win outside of the ways the designer envisioned.”

But on a larger scale, and over a longer timeline, Soros hopes her work will improve the way we interact with technology.

“Thinking more into the future, I want to do work that improves society’s relationship with technology,” she says. “That’s important. A lot of times we develop technology because it’s interesting, exciting, or a good challenge, but how it actually impacts people is sometimes more of an afterthought. Is it positive? Negative? Unfortunately, sometimes we put something into the world and it has a really bad effect. It would be nice if we could predict those things ahead of time or be more careful with safety and impact instead of waiting for bad things to happen. Doing more research that makes things better would be ideal.”


 

For more interviews on evolutionary algorithms and research into evolutionary AI, be sure to check out the interviews in the related resources sections below. And if you’d like future interviews and technical articles delivered direct to your inbox, subscribe to the Lionbridge AI newsletter here.

Subscribe to our newsletter for for more news and interviews:
The Author
Hengtee Lim

Hengtee is a writer with the Lionbridge marketing team. An Australian who now calls Tokyo home, you will often find him crafting short stories in cafes and coffee shops around the city.

    Welcome!

    Sign up to our newsletter for fresh developments from the world of training data. Lionbridge brings you interviews with industry experts, dataset collections and more.