New neural networks coming to PA

does the neural net AI learn on each individual client, or is the learning process still centralized at Uber's office?

This is incredibly interesting, it makes me appreciate human intelligence more, if only because I do not have to bump repeatedly into walls.

Yeah I guess no one is very suprised when a worm brain is pretty stupid. But SCIENCE!

That would be a good idea, a true "adaptive AI". Have you considered a distributed computing model, as in, every client sends collected neural net data back to the server, and from that data you can put the best AI candidates in a deathmatch (or several deathmatches) and then push the winner into the next build and repeat the process.

Do you want Skynet? Because this is how we get Skynet.

I am actually very interested, because that is simulating a living thing, it does not need to be preprogrammed by a human, and it does need to be programmed to learn. You could make several thousand random mutations of that neural data, and then basically artificially evolve the virtual code via virtual natural selection in a short time that would take millennia in nature.
http://www.damninteresting.com/on-the-origin-of-circuits/
this is a good example, except with hardware evolution.

Sometimes you've got to take a long weekend if you ever want to get any work done.

There is a really good documentary you may like, where they made robots capable of learning.

Right when the robot is turn on for example it instantly falls down (its on spider like legs) and it's amazing to watch it learn to stand up, and even walk!

machine learning is an amazing field. There is a lot of work done at my university (university of alberta) in machine learning and robotics.

Just recently they actually release a psycho poker AI that learned how to play just from the base rules. fun stuff http://poker.srv.ualberta.ca/about

I have a few questions about the AI learning process...
You said a while back that the AI plays against itself, right? I am definitely in favor of the AI learning differently on each client (and having those AIs moddable, sharable, able to be reset to default, and the server should, if possible, be able to do the learning for the client, and then send the data back after the game). Some of those are pretty far-fetched, but the AI would definitely be better if it adapted to the player playing against it, especially in galactic war. It'd force the player to be just as strategically flexible in a game vs. the AI as a game vs. the player.

and I forgot the other question I was going to ask :|

I think the problem with adaptive AI is to get it to adapt quickly enough. If I'm not mistaken, it takes many, many games to build the neural net data (if you want to get an idea of what it can take, read the short article I linked in the post above yours). @sorian would be able to give a better idea of the requirements of course though.

I specifically like the automation, the only thing that needs to be be programmed by people is the learning process and an end goal, it opens up many organic solutions to problems that people would never think of. I specifically, would like to see an AI start with the wright brother's aircraft and try to reach the end specifications equivalent to that of a modern passenger jet.

what better place to play many many games than in an Uber sized galactic war.

Hmmm, I'm curious, how much of a diminishing returns curve do you see with training? Would running say 5000 sessions have much of an impact over 1200?

Get ready for the Steam forum to rage at this.