How StarCraft II Can Help Environmentalists Study Life on Earth

It is unlikely that Lou Barbet dares to call himself an avid gamer. He deals with environmental issues at the University of Rennes in France, spending most of his time among plants. But one game has captured his imagination since childhood: StarCraft is a popular online strategy game in which players accumulate resources and create armies of alien fighters to wage wars in extraterrestrial territories. “I’m not a player,” says Barbet, “but I understand what’s going on in the game.”

A few years ago, while playing StarCraft II (the latest version of the game), Barbe realized that besides all the explosions and laser strikes, there was something else going on in the game. He noted that events in StarCraft develop in the same way as any ecosystem develops. “The game has an environment,” says Barbet. “The game has resources and organisms that compete with each other in that environment. This all fits very well with the definition of an ecosystem.”

At this point, Barbet’s idea was limited. But in 2019, DeepMind, Google’s AI research subsidiary Alphabet, pitched an intelligent agent called AlphaStar against the world’s best StarCraft II players. AlphaStar has surpassed 99.8% of human gamers in skill, earning the coveted title of Grandmaster – the highest rank in the game – and thus completing the list of AI victories over humans.

After that, it occurred to Barbe that AlphaStar’s abilities might not be limited to controlling the actions of aliens on a fictional planet. If StarCraft functions like an ecosystem, perhaps game algorithms could help researchers study Earth’s environmental problems?

In an article published in Trends in Ecology and Evolution in 2020, Barbet, along with fellow ecologists from the University of Rennes and Brigham Young University, explains how AlphaStar’s ability to manage StarCraft’s complex multidimensional dynamics can be used to test ideas about ecosystem dynamics in the real world – traditional models cannot yet cope with the solution of this problem.

For example, researchers can run AlphaStar agents on StarCraft maps designed to simulate realistic resource allocations and create models that can provide insight into how different organisms respond to abnormalities such as invasive alien species or habitat destruction.

“The AlphaStar algorithm,” says Barbet, “may have inadvertently become the most complex ecological model in existence.”

The idea of using powerful artificial intelligence tools to analyze environmental problems is not new. Even 15-20 years ago, AI tools were used relatively rarely in ecology, but researchers note that recently there has been a rapid increase in the use of AI in this area – from classifying species of wild animals to predicting an increase in beetle populations in pine forests.

According to ecologists, AI tools, combined with new capabilities for collecting large amounts of data about the Earth, will allow to redefine the methods of studying ecosystems and increase the ability of humans to predict such changes. Sophisticated algorithms like AlphaStar, often designed for purposes unrelated to ecology, can help with these studies.

“Complexity [большинства] ecological models are negligible compared to the complexity of some artificial intelligence systems, says Ben Abbott, an ecologist at Brigham Young University and co-author of an article on AlphaStar. “What we, ecologists, can do, cannot be compared with what these algorithms can do.”

How the champion was created

For AI researchers, StarCraft II, released in 2010, was a challenging intellectual challenge. Just like in chess or go, StarCraft players control different squads to attack opponents, but here players also choose where and when to collect resources, when to create new squads, which squads to create, and so on, all surrounded by many extraneous factors. In chess in one position a player has a choice of about 35 possible moves, in Go – from 200-250. But in StarCraft II there are 1,026 possible moves in one position. Also, unlike games called “full information” games by game theorists, where all players can see the entire game space, StarCraft takes place on a huge map that gamers can see only partially.

An additional complication lies in the fact that gamers play as one of three alien races – terrans, protoss or zerg, each of which has strengths and weaknesses.

DeepMind researchers used machine learning techniques to train AlphaStar’s algorithm and create an AI that can outperform the best players in StarCraft II. They started by creating a league of intelligent agents trained from hundreds of thousands of StarCraft matches played between humans. Then they forced the members of the league of virtual agents to play against each other, selected the strongest of them, made certain adjustments to them, and sent them back to the league. They repeated this process until they gave AlphaStar the power of a juggernaut to destroy everything in its path. Oriol Vinyals, who led the team at DeepMind, the creator of AlphaStar, compared the league itself to an ecosystem in which the process of natural selection operates. “In creating the AlphaStar League, we took inspiration from the literature on evolution,” he says.

The slow-growing Terrans, one of the three alien races in StarCraft II, behave in a one-on-one game like a cactus. — Slow-growing Terrans, one of the three alien races in StarCraft II, behave in a one-on-one game like a cactus.

“Traditional” AI researchers drew inspiration from nature, while Barbe and his fellow environmentalists began to draw inspiration from the game. Published in 2020, they detail the deep parallels between the terran, protoss, and zerg races in StarCraft and the competitive strategies specific to certain organisms.

For example, zerg squads are born colonizers, but weak fighters. They behave like weeds: small, thin, frail, but after disturbing the ecosystem, they are the first to sprout life. Protoss, on the other hand, behave like ferns. They are resource intensive and grow best in groups. Terrans resemble cacti: they grow slowly, but they hold the defense perfectly. As in a real ecosystem, these “species” use their own strategies in the struggle for resources in complex patterns of interaction.

Barbet believes that observing interactions between AlphaStar agents in StarCraft could be a way to test hypotheses about ecological and evolutionary processes that conventional statistical models cannot model, such as predicting how a small change in the available resources in one corner of the StarCraft map will affect troop behavior. terrans and zerg fighting in the opposite corner (though Barbe has not yet tested this idea of his in practice).

Imagine that terrans and zerg are pines and bark beetles, and you will realize that such information can be of considerable value to environmentalists. “For scientists, this game can be a sandbox for experimenting with ecosystems,” says Barbet.

“This creates a very interesting toy model – you see a very simplified system and you can ask very specific questions,” says Ann Tessen, a data scientist at Oregon State University who is not involved in StarCraft’s environmental sense work. that you are dealing with a model. “

Fashion technology

We have to admit that StarCraft II is – for all its complexity – much simpler than a real ecosystem. Barbet notes that the game does not reflect the basic natural processes, for example, the nitrogen cycle, and also does not affect key relationships between organisms, for example, parasitism. Plus, there are only three types of creatures in the world of StarCraft II.

“The problem, in my opinion, is that game mechanics, designed to be as entertaining as possible, only partially reflect the real physical world,” says Werner Rammer, an ecologist at the Technical University of Munich. Rammer argues that it is premature to draw conclusions about AlphaStar’s abilities outside of StarCraft by just watching him play, however complex and intricate it may be.

However, regardless of whether ecologists will ever use AlphaStar in their research, it should be recognized that increasingly sophisticated AI tools are being used to solve environmental and environmental science problems.

A decade ago, Tessen says, the applications of AI in ecology and environmental science were mostly limited to classification tasks, such as quickly identifying species when analyzing recordings of birdsong or landscape types in satellite imagery. Now, he said, AI in ecology goes beyond the narrow scope of object classification and takes on more diverse and ambitious tasks, such as making predictions through the analysis of disordered multidimensional data, that is, exactly the kind of data that ecology usually deals with.

However, AI is still underused in ecology, according to Nicolas Lecomte, a researcher at the Canadian Department of Polar and Boreal Ecology and an ecologist at the University of Moncton in Canada, who uses AI tools to classify the sounds of Arctic birds and predict their migration.

Environmentalists generally do not have the programming skills needed to train artificial intelligence algorithms, he explains. Abbott agrees, adding that gathering enough data to train algorithms is not an easy task. Some data are fairly easy to obtain, such as by analyzing satellite imagery, but other data can be difficult to collect, such as collecting soil samples.

In part, these problems are due to inadequate funding and a shortage of qualified environmentalists, Abbott says. After all, ecology, he notes, is not the most “monetized” area of science.

“Companies like Blizzard, which created StarCraft, spend hundreds of millions of dollars annually developing algorithms for their games,” he says. “They just have a lot more resources than we do. But we, of course, believe that our problems are much more important than their problems. ”Although it was a joke, it is the truth, after all, life on Earth is far from a game.

Machine learning and artificial intelligence continue to penetrate deeper and deeper into the most diverse areas of knowledge, finding more and more non-obvious and non-standard ways for this. If you are interested in new approaches to machine learning and deep learning, a variety of experiments with models, as well as the algorithms underlying the models, you can pay attention to the course “Machine Learning and Deep Learning”, of which NVIDIA is a partner.