id	author	title	date	pages	extension	mime	words	sentences	flesch	summary	cache	txt
cord-195224-7zfq0kxm	Menda, Kunal	Scalable Identification of Partially Observed Systems with Certainty-Equivalent EM	2020-06-20		.txt	text/plain	6803	397	56	Autoregressive approaches directly map a time-history of past inputs to observations, without explicitly reasoning about unobserved states (Billings, 2013) , and are the stateof-the-art approach to the aforementioned problem of modeling the aerobatic helicopter (Punjani & Abbeel, 2015) . Recurrent Neural Networks (RNNs) (BailerJones et al., 1998; Zimmermann & Neuneier, 2000) are a form of black-box non-linear SSM that can be fit to observation and input time-series, and Subspace Identification (SID) methods (Van Overschee & De Moor, 1994) can be used to fit linear SSMs. However, in many cases prior knowledge can be used to specify structured, parametric models of the system (Gupta et al., 2019; in state-space form, commonly refered to as gray-box models. • CE-EM can be faster and more reliable than approaches using particle approximations, • CE-EM scales to high-dimensional problems, and, • CE-EM learns unbiased parameter estimates on deterministic systems with unimodal p(x 1:T | y 1:T ).	./cache/cord-195224-7zfq0kxm.txt	./txt/cord-195224-7zfq0kxm.txt