Alexander Jung
Playing Mario with Deep Reinforcement Learning. Left: Game; right (top): input into the model and area focused by the Spatial Transformer; right (bottom): direct reward and Q-value of the best action.
Code at https://github.com/aleju/mario-ai .
Source
AI playing Super Mario World with Deep Reinforcement Learning
amazing stuff. what's your system configuration and how long it took to train ?
This is really brilliant. And I wonder whether pixels would the best way to define states?
so if i want to try it how do i do it ;( …. (NOOB QUESTION ALERT)
but does it generalize leves? If I were to give it a complete new mario level it has never seen before, would it be able to play? Otherwise its just clever bruteforce of action sequences
просто нереально, машина, не имея представлений об играх, каким-то образом анализирует свои ошибки и понимает, что нужно, чтобы победить!
казалось бы, не такая сложная нейросеть, в которой заложены понятия о итоговом счете и управлении и мотивация победить, но сегодня в ней заложены 3 детали, а завтра возможно настоящего человекоподобного робота соберут, у которого немногим меньше 86 млн нейронов! просто круто
stupid game, no diff smalx bigx, no power ny that, not save the sht, just fk it
This is awesome. You're awesome sir!
That moment when a computer is better at Super Mario World than you are
My pride hurts
As somebody trying to get into Deep Learning, thanks so much for your detailed explanations on your git! That really helps 🙂
OH DEAR GOD THEY ARE EVOLVING HAL IS A DECADE AWAY!
I think it will be fascinating when this type of machine learning is combined with image recognition. For example: "this looks like a pipe – I wonder if I can go down it? This looks like a button – I wonder if I can jump in it?"
Why use pixels for input data rather than the RAM?
Also why are you training it continously on 1-1, wouldn't it be better to have a folder of about 1000 save states and get it to randomly load one? It could recognise water levels and boss battles eventually.
Охуенно