Extracting a stable and compact representation of the environment is crucial
for efficient reinforcement learning in high-dimensional, noisy, and
non-stationary environments. Different categories of information coexist in
such environments -- how to effectively extract and disentangle