You throw open the door to the arcade and come face-to-face with a towering beauty. Your first quarter is spent quickly – she defeats you without much trouble. But with every try, you learn one of her weaknesses and put up more of a fight. To this day, the battle rages on between man and the pinball machine.
Most of us would just accept this rivalry, but computing scientists - forever pushing the limits – are endorsing a new challenger, in the form of an artificially intelligent (AI) pinball wizard. Adam Metcalf, a master’s student in Computing Science decided to take on this project. However, this is by no means an easy task. The AI agent needs to not only track the ball, but it must react in time to operate the flippers as well.
“This is accomplished with a camera mounted above the pinball machine and a hardware controller to fire the flippers,” says Adam. “The framework tracks the ball and sends updates of the ball’s location and estimated velocity to the AI agent. Tracking the ball with the camera was definitely the most challenging task.”
To have a benchmark to measure the AI wizard against, Adam had to play hundreds of games of pinball. The results? Well, with the simple goal of keeping the ball alive for as long as possible, the AI agent is still a novice.
Although the AI agent was able to keep a consistent average time, the human performance increased over time. This shows that the human was learning to play better, something that the AI has not yet been programmed to do.
In addition to the AI, Adam has also constructed a protocol for multi-player pinball games. This allows two players to share a game on different machines at different locations! Working with researchers from the University of Southern California, Adam demonstrated this proof of concept with a cooperative game called Blackout.
“Multi-player pinball games allow us to test the AI agents in new ways,” says Adam, “and give us the ability to construct new games for pinball.”