It's an interesting experiment man, I wouldn't have thought that the AI would have such a hard time playing this game because it's so easy for a human. And if you were to run this experiment for 1-3 days, do you think it could be perfected?
You are viewing a single comment's thread from:
Ideally, the more training it undergoes, the more it learns. But to perfect it I might have to feed it more information like its current distance from the nearest obstacle, maybe the number of obstacles in the scene, its vertical height, maybe a parameter for the changing game speed etc.
That will make it observe a lot of things every step and will improve it a lot. But as it gets complex, it needs more calculations per step and needs much more computational power, which my system doesn't have. Also, all the past experiences have to be stored in memory and for a large model you might need tens of GB of free RAM or VRAM if not 100s of GB.
But in my current state, I can train it for a few more hours and see if it learns more. But I stopped the training because it was getting slow. Also, if it wasn't a web game, you can run 100s of games in parallel and the training time will be much shorter. But again, you need more processing power to run all of those simultaneously.