Reward driven discovery of the optimal microstructure representations with invariant variational autoencoders