Sexy Windsurfing
Fri, Oct 02, 2009 at 18:55:01 Adam Parrish wrote:
Sexy Windsurfing is an automated agent for playing Apples to Apples. Specifically, it’s a program that “emulates” a player in Apples to Apples, using Google search results to determine which cards to play. The agent does a search for the card from its hand (the red card) plus the card under judgment (the green card), and uses those numbers to choose which card to play.
Here’s a screenshot of the interface (text-based, for the moment):
The purpose of the project is to explore a couple of questions: first, is it possible to computationally model how players play Apples to Apples (which is, by all accounts, a highly subjective game)? If so, is there an ideal strategy for players to follow in order to win the game? Most importantly: can an automated agent be fun to play with?
Playtest
I played two games with the automated agent, using the following two strategies: (a) always choose the combination with the highest search results; (b) always choose the combination with the lowest search results. (In a later iteration, I’d like to add random choice as another strategy).
A big thank you to my playtesters!
The computer won two rounds in both games. The “choose fewest hits” variant won with “frazzled science fiction” and “visionary michael jordan”; the “choose most hits” variant won with “timeless lucille ball” and “sappy elephants.” In neither case did the agent win the entire game.
Evaluation
I found that my search algorithms didn’t really approximate the human behaviors that I was trying to emulate. The “choose fewest hits” algorithm, designed to choose more “unusual” pairings, was actually perceived by players as more predictable than the “choose most hits” algorithm. In both cases, players reported that they believed themselves to have a good idea about which card was played by the computer.
Clearly, I need either another corpus or another algorithm. The Google search find two terms when they co-occur on a page, but it doesn’t matter where they occur on the page; I think the algorithms could be greatly improved simply by being able to search for closer collocations of terms.
Players didn’t notice any difference in the level of fun in the game; one player claimed it was “an interesting challenge” to play against the computer.
There were also some methodological issues: because I was the human acting on the agent’s behalf, I sometimes revealed details about how the algorithm was working, or expressed emotions about the cards that were being played (“virtuous rubber gloves” being a particularly amusing pairing). Ideally, I would do a double-blind version of this experiment, where players are isolated in different rooms and don’t know which opponents are computer controlled and which are human.
But that doesn’t seem like a lot of fun.
You can access the github repository for this project here.
Nice meeting you, Adam. Here’s a link to visual notes I created to describe your Apples To Apples project.
http://www.flickr.com/photos/jonnygoldstein/3975013453/in/photostream/
