TL;DR: In POMDPs, the pure exploration task over latent states can be addressed by looking at observations only, and the induced mismatch is far from being hopeless. We show when this is the case and how to simply overcome the possible (structural) limitations under the assumption of knowing at least the observation model. https://arxiv.org/pdf/2406.12795
TL;DR: We know that in POMDPs the pure exploration task over latent states can be addressed by looking at the observations only. Yet, passing through beliefs and solving the task over states sampled from them shows promizing properties (compared to working with observations only). In particular when the latter are not behaving so well or their model is not accessible. https://arxiv.org/pdf/2406.02295
TL;DR: Distributional Policy Evaluation offers new tools to build distributed (a.k.a. disjoint) representations of the state space, if we look for representations that induce the maximum entropic distribution of returns compatible with the returns of a policy. https://proceedings.neurips.cc/paper_files/paper/2023/file/2a98af4fea6a24b73af7b588ca95f755-Paper-Conference.pdf
TL;DR: Including sensory-motor feedbacks in the control policies result in the emergence of motion patterns. In particular, a feedback describing the extent to which a perceived reaction matches the intended action results in an energy efficient policy as well. https://www.frontiersin.org/articles/10.3389/frobt.2021.632804/full
TL;DR: Including sensory-motor feedbacks in the control policies result in the emergence of motion patterns. In particular, a feedback describing the extent to which a perceived reaction matches the intended action results in an energy efficient policy as well. https://infoscience.epfl.ch/record/272142?v=pdf