Monday, October 01, 2007

Serious Gaming For More Human-Like AI

Serious Games simulating the complexity of human interaction

Via: The Restaurant Game Project At MIT Media Lab

Artificial Intelligence, or AI for short, has long been a challenging part of game development. Accurately simulating the complexity of human interaction is precisely what The Restaurant Game, a research project at the MIT Media Lab, is trying to accomplish. By leveraging thousands of actual human interactions within the project, the researchers behind it hope to algorithmically combine these experiences to create a new game with more nuanced, human-like AI.

The Restaurant Game is a multi-player simulation where players take on the role of customers and waitresses. Players are anonymously paired on-line in virtual restaurants where they can chat and interact with the environment.

This game was created with Torque v1.5 from GarageGames

By combining the game play experiences of thousands of players, they shall be able to apply machine learning algorithms to data collected and produce a new single-player game that will enter into the 2008 Independent Games Festival.

Everyone who plays The Restaurant Game will be credited as a Game Designer. All contributions are not equal, however. Designers will be ranked based on how well they play their assigned roles, and accomplish their objectives. There will be only one Lead Designer. Remaining credits will be divided into Game Designers and Assistant Game Designers, and within each category individuals will be ranked
according to the quality of their performance(s). Quality will be determined computationally, based on a number of factors.

The lead developer responsible for the project is Jeff Orkin, a game industry vet who has previously worked on award-winning AI systems for F.E.A.R. and No One Lives Forever 2. Jeff is a PhD candidate in Professor Deb Roy's
Cognitive Machines group at the MIT Media Lab. His research focuses on teaching A.I. characters to communicate and collaborate by observing humans playing online multiplayer games.

To keep the research as accurate as possible, a tutorial must be completed before the player is allowed to play with--or against, depending on how combative your personality is--a real person. The tutorial is relatively short and simple. It gives you instructions on how to order food, pay your bill, and other stuff you probably know how to do if you've ever eaten or spent money. A dialog box pops up whenever the mouse is hovering over an object, and a bunch of possible interactions (such as "pick up" or "eat") with said object are presented. Despite being a tutorial, the player is still provided with the complete freedom to verb any noun in the scene.

When you enter the game itself, a list of available servers is displayed in which two people enter--the first as a waitress, the second as a customer--and interaction ensues. Players can talk to each other by typing out questions, food orders, and bad pick-up lines. After each person has played their part, the customer can walk outside and conclude the demo.

The Restaurant Game takes about 10 minutes to play. You are encouraged to play multiple times. In order for this project to be at all successful, there is a need to collect a lot of data -- data from over 10,000 game play sessions. Play early, play often and please spread the word!

This project attempts to address two frustrations often experienced by professional game developers: 1) Convincing human social behavior is difficult to model with existing hand crafted AI systems. 2) Play testing by people outside of the development team typically comes too late to have a major impact on the final product.

The goal of the Restaurant Game Project is to develop game-engine artificial intelligence that better mimics the behavior of real people. When you play (either as a waitress or a customer) the game learns from what you do, and down the line it will be released as a 1-player game, with game designer credits for all the players. As for the game play now, you interact with another person to either buy or serve food, and afterwards rate the other’s intelligence, kindness, etc., and guess their age (and what they eat for breakfast), which is always fun.