Google often talks about things like “the next billion users,” especially how the products and features it designs will help them get into the company’s ecosystem. Erin Hoffman-John, creative director of Research and Development at Stadia, said in a recent interview withmedia that her team is more focused on “the next billion players.” If Google is going to reach that segment of the user, Hoffman-John says it will start with the tools it provides creators.
“We think that in order for the game to reach a truly broad audience and engage all types of new developers, we have to make game development easier and smaller teams more efficient,” Hoffman-John said.
To that end, Hoffman-John’s team has been working on machine learning to help address some of the common pain points and bottlenecks that developers encounter. The team is made up mainly of game developers and engineers to help apply some of Google’s existing technology to the creative prototype of the game.
Hoffman-John says they see the work as “quite long-term” and even prove that the technology they’re exploring takes two to five years to do, much less than when they’re on the market and for a game release.
She presents a case of a project called Chimera working as a Stadia research and development team, a bit like the idea that one day machine learning tools could allow a 20-person development team to build a game as big and complex as World of Warcraft. But Hoffman-John was quick to admit that it was a bit distant, so her idea was to start with using machine learning to simplify the development of small projects, such as The Card Collection Game (CCG) Magic.
For many true CCGs, Hoffman-John points out, most of the work and budget is done by outsourced artists who draw and design cards. In such strategy games, about 70 percent of development time and investment is spent on repetitive content production, such as creating smaller-differentiated monsters to fill the game world, Hoffman-John said.
“It’s not something that game developers really want to do creatively, and filling a content workflow is something you have to do to enrich the game world,” says Hoffman-John.
So, while working on the Chimera project, the Stadia team wanted machine learning to create these monsters for them. The team took inspiration from the confrontational network used by This Person Does Not Exist, which used a machine learning model trained in live photos to create fake photos.
Hoffman-John says Chimera uses the same rules, and artists create animal models and divide the composition of a CCG game card into rules: scenes illuminated from above, creatures displayed with frames and dynamic poses, and camera angles from below make it look more powerful. They then used a machine learning model to train to recognize high-quality poses, another model could find a landscape with a background for each card, and apply style filters to give them a hand-drawn look.
Together, Chimera can generate dozens of different cards for developers. That’s when artists began to pick out the number of choices presented to them and told the project to create a new card that incorporated those features. This tool can also give them the ability to fine-tune living things, which is undoubtedly necessary.
“If we let the machine stick the animals together, you get what our team calls nightmare fuel,” Hoffman-John said. “They’re terrible, they’re horrible, but that’s not what we want.” If you let the machine do its own thing, it will give you something far beyond the artist’s intentions. So if we just want to make developers more powerful, we have to let them know how to direct artificial intelligence (work for them) very specifically. “
These tools allow developers to adjust the mixing and matching animal parts to form Chimera’s synthetic beasts. They can tell the system to add wings to different parts or remove them to make one part more like a bird, or another to be more like a fish.
Hoffman-John calls this process “talking to the machine” and it’s not the only part of the project that works like this. In addition to the application of machine learning in resource creation, Chimera has also studied its potential in game design, and in addition to using machine learning to help create the visual sonry of cards, Hoffman-John relies on machine learning to inform the game’s own game play mechanisms.
It is not uncommon for competitive games to discover balance problems only after releasing a product to a wide user base. Making a game look more balanced by experienced developers and game testers is one thing, but making it look like it’s one thing for tens of millions of players to make a game look balanced, and for thousands of players, seeking to accumulate an edge on game release day is another.
Machine learning can help you, says Hoffman-John, because it can test a game millions of times using multiple strategies and find strategies that might be more powerful than the designer wants. For Chimera, she deliberately created a problematic play system that could be too powerful or difficult to test, and used machine learning models to help perfect it.
“With this system, I can try something crazy, and then the machine will tell me what’s wrong, how likely it is to win with those abilities, and then we can weaken them back,” Hoffman-John said. Usually, you have to publish it, which makes people unhappy because it makes people feel like something’s wrong with that ability. Because the system is so complex, I can’t predict what will happen.”
This isn’t the only case of machine learning being used in game development, and Ubisoft has talked about using machine learning in 2018 to test For Honor, but it’s in line with the technology that the research and development team has created to simplify the development process.
The most obvious application of the Chimera project is designed for artists and game planners, but Hoffman says the goal of the Stadia research and development team is to find the best way for machine learning to work in all aspects of game development.
“As we go along, we’re trying to refine the principles of machine learning itself,” she said. But in general, when you’re in a situation that requires hundreds of thousands of possibilities, you need some help turning those possibilities into what you want, and that’s what machine learning does. You can think of it as the next step beyond program generation, and in games like No Man’s Space, you see the limits of program generation.”
So, if experiments like the one by the Stadia research and development team have yielded results, and machine learning could even enable small teams to make games that used to require a big corps of research and development teams to make, what would it have to do with the game budget?
“This raises interesting macro-issues for all game development, ” says Hoffman-John. ” I think it may become a reality, but when you give developers more power, it’s very likely that they’ll do more. I’m not sure if the network will cut costs, but they’re going to be used for other things, hoping it’ll make the game better.”
In any case, she stresses that the full impact of machine learning on the industry will not be known for some time, and that the work her team is doing is still in its infancy.
“There’s a lot we don’t know, and I think what’s exciting for us is to give these things to the developers and see how they’re creative,” Hoffman said. Even we don’t know what to do. We just want to give these tools to the developers and see what they want to do with them.”