Particle Swarm Optimization of a Recurrent Neural Network Control for an Underactuated Rotary Crane with Particle Filter Based State Estimation - Developing Concepts in Applied Intelligence - page 51

Particle Swarm Optimization of a Recurrent Neural
Network Control for an Underactuated Rotary Crane
with Particle Filter Based State Estimation

Sam Chau Duong 1 , Hiroshi Kinjo 1 , Eiho Uezato 1 , and Tetsuhiko Yamamoto 2

1

Faculty of Engineering, University of the Ryukyus, Japan
2

Tokushima College of Technology, Japan

Abstract. This paper addresses the control problem of an underactuated rotary
crane system by using a recurrent neural network (RNN) and a particle filter (PF)
based state estimation. The RNN is used as a state feedback controller which is
designed by a constricted particle swarm optimization (PSO). As the study also
considers the problem with assuming that the velocities of the system are not
obtained, PF is utilized to estimate the latent states. Simulations show that the
RNN could provide a superior evolutionary performance and less computational
cost compared to a feedforward NN and that the PF is effective in estimating the
unobserved states.

Keywords: recurrent neural network, particle swarm optimization, nonlinear
control, underactuated system, particle filter, sequential Monte Carlo method.

1 Introduction

The recent immense growth of computing power has allowed several computation meth-
ods applying in broad areas. One of the most successful implementations is the use of
stochastic and intelligent approaches in control engineering [1]. In particular, the ap-
plications of neural networks (NNs) and evolutionary algorithms (EAs) have brought
much success and convenience in both offline and online designs [1] - [3].
Aside from several advantages of NNs, such as good nonlinear processing ability
and robustness with inherently parallel architecture, recurrent NNs (RNNs) have inter-
esting properties with great potential, such as attractor dynamics and internal memory,
and they have shown superiority to feedforward NNs (FNNs). Since a recurrent neu-
ron already has an internal feedback loop, it captures the dynamic response of a system
without external feedback through tapped delays [4], RNNs are thus dynamic mappings
and are more appropriate than FNNs when applied to dynamical systems. Nevertheless,
despite important capabilities, RNNs are much less popular than FNNs because it is
hard to develop a convenient learning algorithm as well as classic gradient-based al-
gorithms are apparently insufficient [5]. In order to overcome the problems of RNNs
training, the last two decades have shown an increasing amount of interest in applying
EAs to construct RNNs (e.g., see [5], [6]). Among the EA family, particle swarm opti-
mization (PSO) [7], [8] is known as one of the latest methods and it has been shown to
have several advantages compared to conventional backpropagation in training FNNs