Learning States Representations in POMDP

TL;DR

This paper introduces a method for learning latent state representations in partially observable Markov decision processes to improve policy learning from limited observations.

Contribution

It presents a novel approach to encode partial observations into a latent space for better policy optimization in POMDPs.

Findings

Latent representations improve policy accuracy in POMDPs.

The method outperforms existing approaches on benchmark tasks.

Efficient learning of states from partial data is demonstrated.

Abstract

We propose to deal with sequential processes where only partial observations are available by learning a latent representation space on which policies may be accurately learned.