Soft Comput (2014) 18:1645–1653
DOI 10.1007/s00500-013-1170-z
METHODOLOGIES AND APPLICATION
Neural-network-based approach to finite-time optimal control
for a class of unknown nonlinear systems
Ruizhuo Song · Wendong Xiao · Qinglai Wei ·
Changyin Sun
Published online: 10 November 2013
© Springer-Verlag Berlin Heidelberg 2013
Abstract This paper proposes a novel finite-time optimal
control method based on input–output data for unknown non-
linear systems using adaptive dynamic programming (ADP)
algorithm. In this method, the single-hidden layer feed-
forward network (SLFN) with extreme learning machine
(ELM) is used to construct the data-based identifier of the
unknown system dynamics. Based on the data-based iden-
tifier, the finite-time optimal control method is established
by ADP algorithm. Two other SLFNs with ELM are used in
ADP method to facilitate the implementation of the iterative
algorithm, which aim to approximate the performance index
function and the optimal control law at each iteration, respec-
tively. A simulation example is provided to demonstrate the
effectiveness of the proposed control s cheme.
Keywords Adaptive dynamic programming ·
Approximate dynamic programming · Unknown nonlinear
systems · Optimal control · Data-based
1 Introduction
The linear optimal control problem with a quadratic cost
function is probably the most well-known control problem
(Duncan et al. 1999; Gabasov et al. 2000), and it can be trans-
Communicated by D. Liu.
R. Song · W. Xiao · C. Sun
School of Automation and Electrical Engineering, University
of Science and Technology Beijing, Beijing 100083, China
Q. Wei (
B
)
The State Key Laboratory of Management and Control for Complex
Systems, Institute of Automation, Chinese Academy of Sciences,
Beijing 100190, China
e-mail: rzsong@126.com
lated into Riccati equation. While the optimal control of non-
linear systems is usually a challenging and difficult problem
(Jin et al. 2012; Zhang et al. 2011e). Furthermore, comparing
with the known system dynamics case, it is more intractable
to solve the optimal control problem of the unknown sys-
tem dynamics. Generally speaking, most actual systems are
nearly far too complex to present the perfect mathemati-
cal models. Whenever no model is available t o design the
system controller nor is easy to produce, a standard way is
resorting to data-based techniques (Guardabassi and Savaresi
2000): (1) on the basis of input-output data, the model of the
unknown system dynamics is identified; (2) on the basis of
the estimated model of the system dynamics, the controller
is designed by model-based design techniques.
It is well known that neural network is an effective tool to
implement intelligent identification based on input–output
data, due to the properties of nonlinearity, adaptivity, self-
learning and fault tolerance (Jagannathan 2006; Yu 2009;
Fernández-Navarro et al. 2013; Richert et al. 2013; Maji et
al. 2013). In which, single-hidden-layer feed-forward neural
network (SLFN) is one of the most useful types (Huang et
al. 2006b). Hornik (1991) proved that if the activation func-
tion is continuous, bounded, and non-constant, then continu-
ous mappings can be approximated by SLFNs with additive
hidden nodes over compact input sets. Leshno et al. (1993)
improved the results of Hornik (1991) and proved that SLFNs
with additive hidden nodes and with a non-polynomial acti-
vation function can approximate any continuous target func-
tions. In Huang et al. (2006b) it is proven in theory that SLFNs
with randomly generated additive and a broad type of acti-
vation functions can universally approximate any continu-
ous target functions in any compact subset of the Euclidean
space. For SLFN training, there are three main approaches:
(1) gradient-descent based, for example back-propagation
(BP) method (Zhang et al. 2008); (2) least square based, for
123