[D] visual input based fighting game ML agent

I'm wondering in my mind about something for a while, How can i create an ml agent for fighting games which is not based on the other player's input, and do it solely based on the visual queues (just like we do with our monkey eyeballs) ?

What type of architecture would this require to so from home? Another computer with a GPU that procceses the input and returns it to the agent? Also with a human time lag. I …

game machinelearning ml