[2212.14511] Can Direct Latent Mannequin Studying Resolve Linear Quadratic Gaussian Management?


Obtain a PDF of the paper titled Can Direct Latent Mannequin Studying Resolve Linear Quadratic Gaussian Management?, by Yi Tian and three different authors

Obtain PDF
HTML (experimental)

Summary:We research the duty of studying state representations from probably high-dimensional observations, with the aim of controlling an unknown partially observable system. We pursue a direct latent mannequin studying strategy, the place a dynamic mannequin in some latent state area is realized by predicting portions immediately associated to planning (e.g., prices) with out reconstructing the observations. Particularly, we give attention to an intuitive cost-driven state illustration studying methodology for fixing Linear Quadratic Gaussian (LQG) management, one of the crucial basic partially observable management issues. As our important outcomes, we set up finite-sample ensures of discovering a near-optimal state illustration operate and a near-optimal controller utilizing the immediately realized latent mannequin. To the most effective of our data, regardless of varied empirical successes, previous to this work it was unclear if such a cost-driven latent mannequin learner enjoys finite-sample ensures. Our work underscores the worth of predicting multi-step prices, an concept that’s key to our principle, and notably additionally an concept that’s identified to be empirically helpful for studying state representations.

Submission historical past

From: Yi Tian [view email]
Fri, 30 Dec 2022 01:42:04 UTC (49 KB)
Wed, 13 Mar 2024 17:44:52 UTC (38 KB)

Supply hyperlink


Please enter your comment!
Please enter your name here