Authors
Takaomi Hirata, Takashi Kuremoto, Masanao Obayashi, Shingo Mabu, Kunikazu
Kobayashi
Corresponding Author
Takaomi Hirata
Available Online 31 March 2018.
DOI
https://doi.org/10.2991/jrnal.2018.4.4.1
Keywords
deep learning, restricted Boltzmann machine, stochastic gradient ascent,
reinforcement learning, error-backpropagation
Abstract
Hinton’s deep auto-encoder (DAE) with multiple restricted Boltzmann machines
(RBMs) is trained by the unsupervised learning of RBMs and fine-tuned by
the supervised learning with error-backpropagation (BP). Kuremoto et al.
proposed a deep belief network (DBN) with RBMs as a time series predictor,
and used the same training methods as DAE. Recently, Hirata et al. proposed
to fine-tune the DBN with a reinforcement learning (RL) algorithm named
“Stochastic Gradient Ascent (SGA)” proposed by Kimura & Kobayashi and
showed the priority to the conventional training method by a benchmark
time series data CATS. In this paper, DBN with SGA is invested its effectiveness
for real time series data. Experiments using atmospheric CO2 concentration,
sunspot number, and Darwin sea level pressures were reported.
Copyright
© 2018, the Authors. Published by ALife Robotics Corp. Ltd.
Open Access
This is an open access article distributed under the CC BY-NC license (http://creativecommons.org/licenses/by-nc/4.0/).