|
Post by Andrew Rose on Feb 26, 2015 0:48:39 GMT -8
Looks like a team at Google are doing GGP in the 1980's video game space. www.bbc.co.uk/news/science-environment-31623427Made harder in that the input is the raw pixels and they have to infer the game rules from the observed behaviour, rather than being given them up-front.
|
|
|
Post by AIFellow on Apr 8, 2015 5:28:15 GMT -8
It is not only Google's thing. In parallel to GGP, there has been GVG-AI (General Video Game AI) Competition: www.gvgai.net/In 2014, it was hosted at CIG Conference in Dortmund.
|
|
|
Post by alandau on Apr 11, 2015 17:39:13 GMT -8
GVG-AI seems much closer to GGP in many ways, in that there's a relatively simple game description language (albeit much more domain-specific), the ability to easily simulate future states (in some versions of the competition), and heavily discretized game states. Major differences from GGP are a focus on fast play (40 ms steps) and widespread use of stochastic behaviors. And, I believe, lack of multiple players.
The part that makes it unappealing to me is the set of restrictions on how to run players. Programs run on their hardware and multithreading is not allowed. This seems to make it quite difficult to use algorithms that may take a while (like propnets), or mix multiple techniques. It's much easier to say "run this expensive thing in thread 1 and use it if it finishes, and in the meantime use this simple technique in thread 2". (Also, I'm not sure how they handle the effects of garbage collection...)
|
|
|
Post by Lars Ericson on Jul 10, 2015 6:20:44 GMT -8
More interesting from a practical perspective would be to crack the GUI protocol for swcpoker.eu so you can write a GDL-based bot that plays OFC/p and occasionally emits chat comments like "Donk!", "Fish!" and "rtlfmao!!!".
|
|