# How does replacing states with latent representations help RL agents?

I have seen many papers using autoencoders to replace images (states) with latent representations. Some of those methods have shown higher rewards using such techniques. However, I do not understand how this helps the RL agent learn better. Perhaps viewing latent representations allows the agent to generalize to novel states more quickly?

Here are 2 papers I have read -

The other main 'problem setting' I have seen images replaced with a latent state is when the authors are looking at planning. The problem with doing any kind of planning is that a model of the transition dynamics, $$p(s' | s, a)$$, is needed. For high dimensional state spaces such as images, this can be very difficult to predict and even relatively small errors will quickly compound so if you use the model to predict multiple time steps into the future the planner is useless because of these compounding errors. I think there is a discussion on this in this paper (certainly there will be references therein that point you in the right direction).