The Center for Data Innovation spoke with Yves Jacquier, executive director of production studio services at Ubisoft Montreal, a video game developer based in Montreal. Jacquier discussed the many ways video games use AI and how video games can advance real-world applications of AI, such as self-driving cars.
This interview has been lightly edited.
Joshua New: When I was a kid, playing video games against real people was always more fun because the AI that controlled other characters just wasn’t as challenging. How far has this kind of AI come over the last 20 years?
Yves Jacquier: We’ve been through three type of challenges with handling AI characters in video games. First, the player never accesses the AI decisions themselves, but only witnesses their results. As such an AI that detects you but does not handle the proper animation to convey the result will be perceived as “poor AI,” whereas it’s more of an animation problem.
Second, games are increasingly complex. Traditional AI deals with techniques such as decision trees to manage all transitions, i.e., “If this happen then do that.” This works well when you have a fixed number of “thises” and “thats” and not too many of them.
Recent breakthroughs in machine learning open new possibilities to create statistical based AI, i.e., “this happens, and I predict I should do that.” This leaves room for more natural behavior in a wider range of situations, and even to emergent behavior as we don’t need to be restricted to a predefined list of possible situations.
Finally AI is meant to be efficient. A game is meant to be fun. So we have to tweak the AI to be competitive or interesting enough without being overwhelming. This also has been and still is a huge challenge.
New: Other than controlling characters, how might a video game use AI that players might not expect?
Jacquier: In open worlds, your avatar interacts with a world. This world has to be coherent and believable. It means that all the systems that run the world simulation have to be consistently linked. If you have a night and day cycle, you expect having vehicles turning their lights on or different type of characters wandering the city during night compared to during day. The player might not interact directly with those systems, however they are fully part of the experience.
Another aspect is online services. As an example, matchmaking uses AI to predict which players are most likely to have fun with you considering your profile, and vice versa. This is very difficult because the level of fun of a multiplayer game depends on technical aspects such as network latency, the experience or level of each player, but also softer elements like playstyle or online behaviour.
We also use a lot of AI in our production processes, mostly to automate the most boring or repetitive tasks, allowing our techno-creative people to concentrate on where they have their best value.
And finally there is an emerging trend: video games can be used in the development of AI to solve real world problems. If we can train an autonomous car in our virtual reproduction of San Francisco, then it might be a good start for real-life vehicles. And the same thing applies in other domains like healthcare.
New: You run a project called La Forge, which Ubisoft launched in 2017, to facilitate partnerships between Ubisoft and academia. Why is there a need for this?
Jacquier: Ubisoft has a long track record of success, and we need to balance the need to be predictive at releasing hits on a regular basis to ensure revenues, while being predictive at being unpredictable in terms of innovation. While incremental innovations can be planned and executed, disruptive innovations are happy incidents that generally arises when there is a crisis.
We decided to create Ubisoft La Forge as a place where we could reproduce the benefits of a crisis in terms of innovation without the drawbacks. Success, procedures, and silos are the enemies of disruptive innovation because they encourage and optimize incremental innovations, but kill the risk.
We thought of La Forge as an ecosystem that is part of Ubisoft, but has to be beneficial for both Ubisoft and our academic partners. Our prototypes are based on the most recent academic papers, helping to inject deep innovations in the way we make games while enabling more academic papers.
New: Has any interesting research come from La Forge yet, or is it too early to tell?
Jacquier: We are two years old but already have many results. We are working on many topics such as procedural animation, bug prediction, automated testing, and preventing online communities from becoming toxic, to name a few. The most interesting part is we discovered that many innovations that we prototype can be used for the video game industry, but are also useful in other sectors.
New: Can you explain why autonomous vehicle researchers might be interested in video games for their work?
Jacquier: Many games can be seen as simulations. “Watch Dogs 2” is a game that takes place in a pretty good replication of the city of San Francisco. You have pedestrians, road signs, and real physical simulation of vehicles, cars, boats, or even drones.
If you have a new idea for an autonomous vehicle, or want to test new algorithms, you will quickly need to deal with the limitations of real-life vehicles of infrastructures, like closed circuits, dealing with captors, and so on. Or you can iterate quickly in a simulation of the real world, and use this as a basis or proof of concept. You also can create specific scenarios to test existing AI. How will a vehicle behave in such situations that involves weather conditions, pedestrians, other vehicles. There is a lot of research literature dealing with transferring what AI learned in one context into another context, and because games are more and more rich simulations of worlds, we feel that they can now enrich more than only the life of players.