1D State Space instead of RBG Images
OUStudent opened this issue · comments
I've developed an algorithm for solving Reinforcement learning type problems based off using a genetic algorithm to train a population of neural networks. The only problem is that it does not work using convolution layers to process three channel RGB images.
As of right now, this project only returns the RBG images for each action, so trying to process a 240x256x3 without using a CNN is scratched. I thought about reducing the RBG image to that of grayscale, but still the input layer of my network would be 240x256=61440 nodes, to even make sense of that I would have to make my networks 3 layers deep with 100,000 nodes each; which is way to computationally heavy.
Is there any way you could implement an option so that it would have the ability to return a grid around the player that is simplified to 0 for nothing, 1 for object, 2 for enemy, etc.. ? In this way, the state space is only 10x10 or so?
There are no plans to support this type of functionality. Feel free to fork the code and work out a solution however.