Project Galena
Improving gaming with imitation learning
In project Galena, we’re asking how imitation learning can make gaming better.
Imitation learning starts from demonstrations of how to do a task, or from a teacher who knows how to do it. By analyzing the demonstrations or by asking for examples from the teacher, we train an AI agent to do the same thing. The key benefit of imitation learning is that it lets us easily design and edit complex behaviors for AI agents: we just have to show the agent a few examples of what to do, instead of trying to tell the agent what to do by writing code or scripts.
We’re currently using imitation learning to help reduce the effects of lag in cloud gaming. When a player’s connection is poor, it can take too long for actions, like joystick movements and button presses, to reach the game server, leading to a poor gaming experience. We are training an AI agent that can smooth over occasional lag problems by interpolating actions that aren’t received in time.
We’re also looking at other uses of imitation learning. For example, imitation learning can help game designers develop believable non-player characters–imitation makes it easier to understand and edit complex non-player character (NPC) behaviors.