Open ctrembla opened 6 years ago
If it doesn't know any controls or where the score is a generic neural network will take a long time to work out how to do anything. If there's a standard game API on one end it could be doable like a default controller set for a game console and a standard score variable to read for training.... But then there's deciding what to input and the minimum number of inputs is better the only way to be truly generalised is each pixel getting it's own input x3 for RGB channels and if that's the input you need to add layers upon layers for the image recognition and abstraction before even training the game again a standard API could help but there are only so many games which can be coded so there are exactly N inputs. Could be worth thinking about for a lesson in API design though.
For example a number of games might be able to be broken down like this: Player location, player velocity, nearest killing thing x4 (one for each direction NESW), and their velocities, nearest point scoring thing x4 (NESW), level exit direction, score. Which is 39 inputs a lot of data for a toy NN some can default to 0 if unnecessary at a given time player death should tell the NN to iterate and reset the game.
interesting. I was thinking of the mario one I saw and he said his neuro network knew nothing about the game and didnt even know how to press the buttons. Then after a bunch of generations it learned and finally won the game. I've seen some that can play like hundreds of games without being altered.
I think it be cool, to see a neuronetwork that is its own app and can be opened and play any other game. Like the neuro network would know nothing about the game not even the buttons & it has to learn all that stuff, before it can even start playing.