UBC computer science professor brings machine learning to video games
It’s finally happened, folks—video games are the new movies. Remember how hyped everyone got when Kevin Spacey was in Call of Duty? Or when Ellen Page saddled up for Beyond: Two Souls? With their long and involved plots—and the chance to manipulate them yourself—modern games have rightly seen more and more stars crossing over into the realm of consoles.
Source: Kate Wilson
Signing on a big name, however, comes with some drawbacks. Typically, working with big stars requires numerous motion capture cameras, a clear schedule for the actors, and a very large chequebook. But—according to UBC computer science professor Michiel van de Panne—it doesn’t have to be that way.
Rather than drawing numerous tiny dots on actors’ faces, decking them out in suits that detect every motion point, and setting up miles of green screen fabric, van de Panne has devised an algorithm for the computer characters to teach themselves complex motor skills like walking and running. That’s right. Teach. Themselves.
“We’re creating physically-simulated humans that learn to move with skill and agility through their surroundings,” says van de Panne. “We’re instructing computer characters to learn to respond to their environment without having to hand-code the required strategies, such as how to maintain balance or plan a path through moving obstacles. Instead, these behaviors can be learned.”
Trial and error is the basis for the machine learning, with computer characters—and, the researcher hopes, future robots—operating through a type of “reinforcement learning”. In van de Panne’s programming, the character undergoes numerous trial and error experiments in order to attain a reward when it successfully completes a task. Over time, the system progressively identifies better actions to take in given situations.
The algorithm, called DeepLoco, allows characters to move around in a way that is both realistic and sensitive to its surroundings. Simulated characters have so far learned how to walk along a narrow path without falling, avoid any obstacles or people—even if they are moving—and to display complex motor skills like dribbing a soccer ball towards a goal. Watch out, Whitecaps.
“The machine learning is like teaching yourself a new sport,” van de Panne says. “Until you try it, you don’t know what you need to pay attention to. If you’re learning to snowboard, you may not know that you need to distribute your weight in a particular way between your toes and heels. These are strategies that are best learned, as they are very difficult to code or design in any other way.”
Right now, it’s hard to tell how much the algorithm’s learning process mirrors a human’s—just take a look at how Facebook’s robots are inventing their own language, for instance, to see how machines and people differ in their logical processing. But whichever route the simulated characters have taken to achieve their bipedal success, you can be sure they’ll only keep improving.
Which is great news for the gaming industry—and less good for Kevin Spacey’s bank balance.