February 23, 2007
It’s Always 5pm Somewhere
The pilot tests of The Restaurant Game continue! This weekend on Saturday and Sunday we will run tests at 5pm in every time zone. Choose a time zone that works for you, and come online. This should make it a little easier for our friends in Europe (and elsewhere) to participate. As always, you are welcome to play any other time too, but you are more likely to find other players at 5pm.
Download the latest version of the game here:
Read more about this project in my previous post: 1,000 Players Can’t Be Wrong.
Thanks everyone for all of the support in the tests so far. 18 games have been logged in the first two nights, played by 24 unique players. So far the game is only advertised on this blog. Next week I will advertise the game more widely.
How long do you think it will take to get to 1,000 games? “Never” is an acceptable answer. I think we can do it within one month -- by March 25.
EDIT: To answer my own question, two weeks. In two weeks 1,074 games have been played. I am now shooting for 10,000.
Posted by jorkin at February 23, 2007 05:56 AM
Sounds really cool. Blogged at GTxA, hopefully that will help get you a few more participants this weekend.
Downloading the game now. I'll play it when I get a chance.
Maybe I misunderstand what you're saying, but this kind of tech has been developed/used by other game developers (Sony) and middleware companies (Artificial Contender and AILive) so it would be incorrect to say The Restaurant Game is the first game using this approach. The two most popular names I've seen for the technology are: "behaviour capture" and "instance based learning". It's a very interesting field, which I've been meaning to research for some time now (but still haven't found the time to do so!).
Apologies if I've misunderstood you or I'm telling you stuff you already know, but I thought it worth mentioning.
Mat, you are absolutely correct that others have already implemented behavior capture. However, I see this as a bit different from "collaborative authoring" where we combine both the behavior and language (chat text) from thousands of possible paths through a simple restaurant narrative -- a conversational game -- to teach an AI character to play a role (e.g. customer or waitress), and to communicate naturally.
My understanding of the existing middleware is that it learns tactics for "winning" in combat or sports. There is no winner in The Restaurant Game, and no goal to optimize for. Two characters play roles that would take a human author a huge amount of time to hand-script, when accounting for the variation in how and when human players say and do different things in this scenario.
I'm interested in tech that will bring us AI characters that create a rich gameplay experience, rather than tech that brings us characters that are better at beating the pulp out of us. :)
One of the things that players really seemed to respond to in FEAR was the way the enemies communicate with other squad members. This took a huge amount of manual labor and iteration to get right, and in this case the communication was only between AI (and not with human players). I'd like a system that can teach AI characters to handle these social interactions by observing human social interactions.
I guess the point is this: when I finish collecting data with The Restaurant Game, do you think we can feed this data into Artificial Contender or AILive to produce characters that can play as the waitress or customer? If not, there's still something missing from the story of "behavior capture", which is a void I'd like to work towards filling with collaborative authoring.