Wednesday, March 14, 2007

The Restaurant Game Project

"The Restaurant Game is a research project at the MIT Media Lab that will algorithmically combine the gameplay experiences of thousands of players to create a new game. In a few months, we will apply machine learning algorithms to data collected through the multiplayer Restaurant Game, and produce a new single-player game that we will enter into the 2008 Independent Games Festival."

"The primary responsibility of a game AI programmer is to anticipate player behavior, and provide satisfying responses for every anticipated scenario. With each generation of games this gets more difficult, as virtual worlds become more realistic with additions like real-time physics simulation.

I’m working on a research project that flips the traditional approach to game AI on its head. My goal is to capture gameplay from 1,000 multiplayer game sessions, analyze this data, and use it to generate AI for a new single player game this summer.

The simple game I’ve created does not have real-time physics, but it does require conversation between players via typed text, which proves equally challenging -- in both cases no matter how many games you observe, you will never see the exact same scenario play out twice.

It is unrealistic to believe human AI programmers will ever account for the nuances of every possible scenario. Think about an analogy to animation – when animators want convincing human movement, they go straight to the source and use motion capture.

Animators blend, layer, sequence captured animations to generate new motions beyond those recorded. Why should we believe that authoring AI behaviors is any different – perhaps we can get the most believable results by starting with the source material, and procedurally blending, layering, and sequencing behaviors as we see fit."

"The Restaurant Game takes about 10 minutes to play. It is a two-player game that will automatically find partners for players once you join a server. You are welcome encouraged to play multiple times. In order for this project to be at all successful, we will need to collect a lot of data -- data from over 1,000 10,000 gameplay sessions. "

No comments: