I am trying to use a neural network to predict the next state output given the current state and action pairs. Both input and outputs are continuous variables. Due to the high dimensionality of each input, ( ~50 dimensional input ) and 48 dimensional output, I am not able to achieve an achieve a satisfiable enough accuracy.
I am thinking of using an auto-encoder to learn a latent representation of the state. Would a latent representation from an auto-encoder help to improve the prediction accuracy ? and can the latent representation have a higher dimensional space compared to the original state ?