Neural networks in GC3 Game AI?

I was just wondering if anyone at Stardock is currently following the Google DeepMind Go challenge, where an AI from a Google research division is using deep neural networks to beat professional players at Go, something that was considered beyond a computer's ability until quite recently. They got published in the "Nature" journal for describing their approach to use self-learning software to potentially teach a program to solve any problem, and applied it to Go most recently. There is now talk about the company trying their hand at Starcraft or other computer games next, since they already solved several arcade games in 2014. Are there any plans by Stardock to develop their AI by using neural networks yet, or asking Google to try their hand at Galactic Civilisations? This could potentially give the game a "hands-off" and still close to unbeatable AI, if desired... 

28,305 views 7 replies
Reply #2 Top

I think if they were serious about it they'd provide some sort of networked AI based on an server farm.

 

Also, an more interesting question is, can they use AI to DESIGN games? One of SD's affiliates is already on the case from what I've seen...

Reply #3 Top

For the championship games, they run their neural network on a cluster, with over 1000 CPUs and and a couple hundred GPUs. Obviously this wouldn't be practicable for a computer game AI, however they do it mainly since they use several neural networks in conjunction with a tree search. A single neural net could be run with far fewer resources, including a regular desktop or even laptop computer, although training the network sufficiently to actually become strong would probably require a cluster and significant time.

Reply #4 Top

What about some sort of internet neural network that feeds of the Internet's resources.

Reply #5 Top

Skynet? 

Reply #6 Top

Quoting thimon2k, reply 5

Skynet? 
End of thimon2k's quote

My thoughts exactly :)

Reply #7 Top

I figure you just take an expert level player, lock him up in a room with an Internet connection, and make him simulate being a GalCiv AI opponent.  The Turing Test works both ways, you know.  That is going to get us a neural net to play against a lot sooner than waiting for AlphaGo to learn GalCiv for us.  Or Skynet.