servebolt
Awario
Discover

Not only did AlphaStar artificial intelligence beat professional StarCraft II gamers, but it also unintentionally had expertise in ecology.

Lou Barbe doesn’t consider himself a real gamer. As an ecologist at the University of Rennes, France, he spends most of his time with plants. But there is one game that has captured his affection since he was a child: StarCraft, the famous series of strategy games in which players accumulate resources and build alien legions to fight on zones. outer space. “I’m not a good player”Barbe said, “But enough for me to understand what’s going on in the game.”

While playing StarCraft II – the newest version of the series – a few years ago, Barbe realized that, apart from the explosions or the laser beam, something else was going on. StarCraft has something of an ecosystem. Barbe says: “We have an environment. We have resources. We have organisms that are competing for the environment. That is the basic definition of an ecosystem. ”

Barbe then temporarily dismissed this idea. Then comes 2019, when DeepMind, Alphabet’s artificial intelligence research company, comes up with an AI called AlphaStar and has it play against the top StarCraft II players around the world. AlphaStar outperformed 99.8% of players, achieved Grandmaster rank – the highest rating in the game, and also added another victory to the AI ​​”title room”.

And then, Barbe realized that, AlphaStar’s power can go beyond manipulating aliens on a fictional star. If StarCraft behaves like an ecosystem, maybe the algorithms for gaming would be helpful in studying ecological problems on Earth.

Write in a magazine Trends in Ecology and Evolution in 2020, Barbe, along with other ecologists from the University of Rennes and Brigham Young University, explained how StarCraft’s ability to manage complex, multidimensional functions could be transformed to test these The idea of ​​ecological models in the real world is very difficult with traditional models. For example, ecologists were able to deploy AlphaStar on StarCraft maps that were designed for real resource distribution, to model how different organisms react to disturbances such as species. invasion or loss of habitat.

AlphaStar’s algorithm, according to Barbe, may have unintentionally become the most sophisticated ecological model ever.

This idea is part of a recent trend in ecology to use powerful artificial intelligence tools to analyze environmental problems. Although not as popular in the past 15 or 20 years, scientists say industry applications of AI are exploding in recent times, from differentiation of wild species to prediction of ladybug proliferation. in pine forests. Biologists think AI tools, combined with their ability to collect large amounts of data about the Earth, could modify the way ecosystems are studied and increase their ability to predict how they change. Complex algorithms like AlphaStar – often developed with an original purpose unrelated to ecology – can help such studies go far.

“Most ecological models are too small compared to the complexity of some artificial intelligence systems.”Said Ben Abbott, ecologist from Brigham Young University, co-author of the study. “We’re actually just taking the tip of what this approach can offer.”

Build a champion

For AI researchers, StarCraft II has presented a formidable challenge since it was launched in 2010. Like chess or Go, StarCraft players control different units to attack opponents. But they also choose where and when to gather resources, when to create new units, or choose which units for themselves, and many other complications. In comparison, when a turn in chess has about 35 possible moves, and Go is between 200 and 250 possibilities, for StarCraft II, that number is 10 ^ 26 (100 million billion possibilities. ). In addition, unlike “full information” games, where the player can see the entire game space, StarCraft is played on a large map, where each player can only observe one. part. Not stopping there, the complexity of the game also increases (with both humans and AI) when the player controls 1 of 3 alien races – Terran, Protoss or Zerg, with strengths and weaknesses. private.

Not only did AlphaStar artificial intelligence beat professional StarCraft II gamers but it also unintentionally had expertise in ecology | Discover

To create an AI capable of beating the best StarCraft II players, the researchers at DeepMind used machine learning techniques to train AlphaStar. First, they created a group of AI trained (train) using data from hundreds of thousands of StarCraft matches between humans. They then let this group of AI fight each other, choose the best “ones”, shuffle them up a bit, and then let them continue. They repeated this process over and over again until the dominant AlphaStar emerged. Oriol Vinyals, team leader who created AlphaStar at DeepMind, compares this small tournament to a kind of naturally selective ecosystem. “We took a lot of inspiration to design the AlphaStar tournament from evolution texts.”he said.

Since AI researchers were inspired by nature, Barbe and ecologists got their inspiration from the game. In their 2020 study, they delved into the coexistence of the Terran, Protoss, and Zerg races in StarCraft and the competitive tactics of each organism. Zerg units, for example, were agile settlers, but poorly fought, similar to the first plants that grew on ecologically affected areas. Protoss, on the other hand, resemble ferns, use a lot of resources and grow best in groups. Terrans are like cacti: they grow slowly, but they defend well. As a true ecosystem, these “species” use different tactics to scramble for resources in different layers of interaction.

Although he has not officially experimented, Barbe suggests observing the interactions between AlphaStar bots in StarCraft could be a way of testing theories about ecological and evolutionary processes. Conventional statistical models cannot be done. Perhaps, for example, predicting how a small resource change in one corner of the StarCraft map would affect the Terran and Zerg team competing in the opposite corner. Replace the Terran and the Zerg with the pines and beetles, and you can begin to see why a prediction like this is so valuable to environmental managers. “It could be a closed environment” for scientists to experiment with ecosystems, Barbe said.

Anne Thessen of Oregon State University, who is not involved in StarCraft ecosystem research, said: “It can transform into a very interesting kind of experimental model, where you have a simplified system and ask specific questions. Just keep in mind, it’s still an emulator. ”

Technology trends

Obviously, no matter how complex, StarCraft II is still too simple compared to a true ecosystem. Barbe also noted that basic natural processes like the nitrogen cycle do not occur in this game, or important relationships between organisms, such as parasites. And there are only 3 species in the game.

Not only did AlphaStar artificial intelligence beat professional StarCraft II gamers but it also unintentionally had expertise in ecology | Discover

Terrans have features comparable to the cactus from StarCraft II

Mr. Werner Rammer, ecologist at Technical University of Munich commented: “One problem is that, in my opinion, the mechanics of the game – which are designed to be as entertaining as possible – are only similar to the real world in appearance.”. This would make it more difficult to generalize the observations from AlphaStar beyond the game framework, Rammer said.

Whether or not ecologists ever actually use AlphaStar for research, there are more and more sophisticated AI tools being used to solve problems in ecological and environmental sciences.

Ten years ago, according to Thessen, the application of artificial intelligence in the ecological and environmental sciences was limited to tasks such as classification, for example, identifying birds through song recordings or distinctions. landscape type via satellite image. Now, the application of AI in ecology has passed that level, and can perform tasks like forecasting based on messy and multidimensional data – the most common data in ecology.

But according to Nicolas Lecomte, an ecologist from Moncton University in Canada, who uses AI tools to recognize the calls of birds in the Arctic to predict migration patterns, AI has yet to be exhausted. maximum use in this industry. Ecologists may be afraid of the amount of programming skills it takes to train an artificial intelligence algorithm, he explains. And collecting enough data to train the algorithms can also be difficult. Some data, such as satellite imagery, can be easily obtained, but other information such as acres of land is difficult to obtain.

It must also take into account the economic aspect as well as the number of skilled collaborators willing to work in ecology, as he points out, Abbott, as he points out, is not easy research to make money. Companies like Blizzard, the creator of StarCraft, are spending “Hundreds of millions of dollars annually to develop algorithms in our games”he shared. “They have a lot more resources than we do. But we, of course, think our questions are much more important than theirs. “. He is probably only half joking – because from a certain perspective life on Earth is, after all, more than just a game.

Content Protection by DMCA.com
Back to top button