Author Topic: Neural network AIs?  (Read 4866 times)

Offline Chthon

  • Sr. Member Mark II
  • ****
  • Posts: 398
Neural network AIs?
« on: January 23, 2018, 09:06:17 pm »
Hey, Chris, have you thought about using antagonistic neural networks to create an AI that isn't omniscient, but instead it guesses at where your ships are based on what scouting it's done? I've learned about similar ones that take image features, and then reproduce what it thinks the original image looked like.

You could feed this ai similar features as in what ships it's scouted, how many ships of each kind you think the player has, and how many were killed as well as how long ago, and it could guess at where the rest of the player's fleets are.

I feel a learning AI of this type might throw some players through loops, and at the same time make scouting a priority for the AI as well.

(Split by Chris to remove the unrelated quoted text and move this into its own topic)
« Last Edit: January 23, 2018, 09:17:50 pm by x4000 »

Offline x4000

  • Chris McElligott Park, Arcen Founder and Lead Dev
  • Arcen Staff
  • Zenith Council Member Mark III
  • *****
  • Posts: 31,651
Re: Neural network AIs?
« Reply #1 on: January 23, 2018, 09:16:47 pm »
Hey, Chris, have you thought about using antagonistic neural networks to create an AI that isn't omniscient, but instead it guesses at where your ships are based on what scouting it's done? I've learned about similar ones that take image features, and then reproduce what it thinks the original image looked like.

You could feed this ai similar features as in what ships it's scouted, how many ships of each kind you think the player has, and how many were killed as well as how long ago, and it could guess at where the rest of the player's fleets are.

I feel a learning AI of this type might throw some players through loops, and at the same time make scouting a priority for the AI as well.

Well, the cool thing is that the AI is all open-source, so anyone is free to add this sort of thing as a mod.  And if it works well and people prefer it, we can bake it right on into the official builds.  As an AI type or otherwise.

As it stands, right now our priority is polishing what we have and reducing our workload, not increasing it if we can avoid it.  We're already massively overbudget and over schedule, and we'd like for this to release in as pristine a state as we can.  Keeping away from scope creep at this point is the big lesson for us in terms of trying for any kind of cleanness at 1.0.

But the good news is that we've built an overall engine that is so darn flexible that we can start contemplating things like this post-1.0 if the game is selling well enough; and if it's not, modders and similar can contemplate it if they would like to.
Have ideas or bug reports for one of our games?  Mantis for Suggestions and Bug Reports. Thanks for helping to make our games better!