They would essentially train the neural network and once it reaches a satisfactory level of competence they will just save the neural network's neuron values to disk. Then when you want to play against it you just create a neural network and give the neurons the stores values, usually just simple floats in the range of 0 to 1. That is assuming they don't want it to learn continuously after shipping.
Keep in mind the current implementation of this is using some sort of custom bf1 client per agent, i doubt this would be efficient to host, especially if the hosting machine lacks a dedicated graphics card. Maybe it's consolidated into a single client for all agents, but it doesn't matter, you need to render the game in some form for the AI to function.
At the moment the AI is being used for QA, like stress tests or testing crashes.
Did you work on this wranglor? I've been playing around with RL for financial purposes and would love to chat a little bit about it. Let me know how to reach you.
I'd rather have the AI works in a similar fashion to the Director in L4D2, or the Commander in Natural Selection.
Someone should make a loop of those two soldiers spinning around eachother.
is no one worried about the possible implications???
there's going to be knocking on DICE's door with a hefty government contract but little do they know, they've already strapped gopros on thousands of bots, armed and ready to "play the objective"
Sorry, you need to Log In to post a reply to this thread.
