[1G4-OS-21a] 世界モデルと知能

[1G4-OS-21a-03] Basic Consideration and Efficency of Reinforcement Learning on Latent Space of Variational Auto Encoder

Consideration of Reinforcement Learning on Latent Space

〇Masato Nakai¹ (1. Doctoral Program in Intelligent and Mechanical Interaction Systems, University of Tsukuba, Japan)

Keywords:VAE, Reinforcement Learning, Riemannian Metric, Latent Space

Reinforcement learning on images directly requires a large number of training images, but it has been known that using low-dimensional abstract representation is more efficient to perform reinforcement learning. A typical low-dimensional abstract representation is the latent space of Variational Auto Encoder (VAE). However the relationship between the image and the latent variable is not clear because deep learning is interposed between them. Therfore the reason why reinforcement learning is possible by sampling the latent variable could not be clarified. In this paper, we clarify the reason that reinforcement learning is possible by sampling on the latent space of VAE, and improve the efficiency of reinforcement learning based on the reason.