General Game Playing Platforms

For several years, a community of people generating and competing with AI to play games has existed. In recent years this community has expanded considerably, mainly due to the projects at the Stanford Logic Group, so I wanted to give a summary of the field as it stands now.

General Game Playing (GGP) is the art of making AI that can play a diverse range of games well without being hand-programmed for any specific game. Rules are described in a formal language called GDL and agents parse the game descriptions, infer the valid moves at any given time and apply a strategy to win.

There are multiple Game Manager hosts,  which developers can register their agents with. The main two ones are Dresden GGP Server and Tiltyard which provide large game databases played by registered agents. Matches are automatically hosted between agents, across the Internet, and participants are awarded scores and ratings. The time between an agent getting to know the rules and having to play the game is typically less than a minute, so the AI must be truly general-purpose to play well across the large variety of games.

Every year a world cup with prizes is held for the very best agents. Fluxplayer, Ary and Turtleplayer have been strong in the past, but Cadiaplayer from Reykjavik University  has won most world cups and is also the current champion.

The slides from a previous course at Dresden University are a good introduction on how to get started. Several existing frameworks and reference players help to get up to speed quickly. There is also an online  course starting in September at Coursera that will teach all the relevant theory and practical knowledge to participate.

My next step is to put together a framework that can parse GDL and communicate to the Game Manager via the established protocol, and then start plugging in my previous AI algorithms. Exciting times ahead! :)