Nonlinear Modeling of Dynamic Systems with

the Self-Organizing Map

Guilherme de A. Barreto and Aluizio F.R. Araújo

Department of Electrical Engineering, University of São Paulo (USP)

Av. Trabalhador Sancarlense, 400, CEP. 13560-590, São Carlos-SP, Brazil
{gbarreto, aluizioa}@sel.eesc.sc.usp.br
http://www.sel.eesc.sc.usp.br/lasi.htm

Abstract. In this paper we propose an unsupervised neural modeling

technique, called Vector-Quantized Temporal Associative Memory (VQ-
TAM). Using VQTAM, the Kohonen’s self-organizing map (SOM) be-
comes capable of approximating dynamical nonlinear mappings from
time series of measured input-output data. The SOM produces mod-
eling results as accurate as those produced by multilayer perceptron
(MLP) networks, and better than those produced by radial basis func-
tions (RBF) networks, both the MLP and the RBF based on supervised
training. In addition, the SOM is less sensitive to weight initialization
than MLP networks. The three networks are evaluated through simula-
tions and compared with the linear ARX model in the forward modeling
of a hydraulic actuator.

1 Introduction

Training a neural network using input-output data from a nonlinear plant can
be considered as a nonlinear function approximation problem. In this paper,
it is proposed a new neural modeling technique which uses unsupervised neu-
ral networks for function approximation, instead of the usual supervised ones
(MLP and RBF) [1]. By means of this technique, called Vector-Quantized Tem-
poral Associative Memory (VQTAM), it is shown that the Self-Organizing Map
(SOM) [2] can be successfully used to approximate nonlinear input-output map-
pings. Computer simulations illustrate this approximation ability of the SOM
using the VQTAM approach and compare the obtained results with those pro-
duced by MLP and RBF networks and linear models.
The remainder of the paper is organized as follows. Section 2 introduces the
VQTAM technique and its main properties. It is also shown how the VQTAM
can be used together with SOM in the identification of forward dynamics of a
hydraulic actuator. In Section 3, the approximation results of the SOM algorithm
are compared with those produced by other linear and nonlinear methods. The
paper is concluded in Section 4.

J.R. Dorronsoro (Ed.): ICANN 2002, LNCS 2415, pp. 975–980, 2002.

c Springer-Verlag Berlin Heidelberg 2002
976 G. de A. Barreto and A.F.R. Araújo

2 Temporal Associative Memory Using the SOM

The Self-Organizing Map (SOM) is an unsupervised neural algorithm designed
to represent neighborhood (spatial) relationships among vectors of an unlabelled
data set [2]. The neurons in the SOM are put together in an output layer, A, in
one-, two- or even three-dimensional arrays. Each neuron i ∈ A has an weight
vector wi ∈ n with the same dimension as the input vector x ∈ n . An impor-
tant characteristic of the SOM algorithm is that it can learn only static input-
output mappings [3], which are described mathematically by y(t) = f (u(t)),
where u(t) ∈ n and y(t) ∈ m denote the input and output vectors, respec-
tively. This paper however is interested in showing that, with slight modifica-
tions, the SOM can also be used to approximate nonlinear dynamic mappings,
such as those described as follows [4]:
y(t + 1) = f [y(t), . . . , y(t − ny + 1); u(t), . . . , u(t − nu + 1)] (1)
where nu and ny are the maximum delays (orders) of the input and output
memories, respectively. For the SOM to be able to learn dynamic mappings,
it must have some type of short-term memory (STM) mechanism [5]. That is,
the SOM should be capable of temporarily storing past information about the
system’s input and output vectors.
By using STM mechanisms, such as delay lines, the SOM algorithm is used to
approximate the nonlinear function f (·). To achieve this, the input vector to the
SOM, x(t), is slightly modified so that it has from now on two parts. The first
part, denoted as xin (t), carries data about the input of the dynamic mapping
being learned. The second part, denoted xout (t), contains data related to the
output of this mapping. The weight vector x(t) of neuron i has its dimension
increased accordingly. These changes in x(t) and wi (t) are written as:

in 
in 
x (t) wi (t)
x(t) = and wi (t) = (2)
xout (t) wiout (t)
It is worth noting the difference between this unsupervised strategy and that
used in training supervised networks. In MLP and RBF networks, the vector
xin (t) is presented to the network input, while the xout (t) is used at the network
output to compute the error signal that guides network training. When using the
SOM with the definitions in (2), the vector xout (t) is presented to the network
input together with the vector xin (t) and no error signal is explicitly computed.
Depending on the variables chosen to build the vectors xin (t) and xout (t) one
can use the SOM algorithm to learn the forward or the inverse dynamics of a
nonlinear plant. For example, if one wishes to approximate the forward dynamics
as defined in (1) the following definitions apply:
xin (t) = [y(t), . . . , y(t − ny + 1); u(t), . . . , u(t − nu + 1)] (3)
xout (t) = y(t + 1) (4)
During training, the winning neurons are found using only the portion corre-
sponding to xin (t):
i∗ (t) = arg min{xin (t) − wiin } (5)
i∈A
 Nonlinear Modeling of Dynamic Systems with the Self-Organizing Map 977

In updating the weights both, xin (t) and xout (t), are used:

∆wiin (t) = η(t)h(i∗ , i; t)[xin (t) − wiin (t)] (6)

∆wiout (t) = η(t)h(i∗ , i; t)[xout (t) − wiout (t)] (7)

where η(t) is the learning rate andh(i∗ , i; t) is a Gaussian-type

 neighborhood
function given by h(i∗ , i; t) = exp −ri (t) − ri∗ (t)2 /σ 2 (t) , where ri (t) and
ri∗ (t) are, respectively, the locations of neurons i and i∗ in the output ar-
ray. The variables η(t) and σ(t) decay exponentially with time according to
(t/T ) (t/T )
η(t) = η0 (ηT /η0 ) and σ(t) = σ0 (σT /σ0 ) , where η0 and σ0 denote the
initial values of η(t) and σ(t), while ηT and σT are the final ones, after T train-
ing iterations. As training proceeds, the SOM algorithm learns to associate the
outputs, xout (t), of the dynamic mapping being learned with the correspond-
ing inputs, xin (t). At the same time, the SOM performs vector quantization on
both the input and output spaces of the mapping. Thus, this technique will be
referred to as Vector-Quantized Temporal Associative Memory (VQTAM).
For validation, the vector y(t + 1) is unknown, and hence, the vector xout (t)
is not built. A trained SOM is then used to obtain estimates for this output
vector from the output portion of the weight vector, wiout (t). For the case of
estimating the forward dynamics, one gets:

ŷ(t + 1) ≡ wiout
∗ (t) (8)

where the winning neuron, i∗ (t), is found as defined in (5). The estimation
process continues for M steps until an entirely new time series is built from the
estimated values of y(t + 1).

3 Computer Simulations

Figure 1 shows measured values of the valve size (input variable), u ∈ , and
the oil pressure (output variable), y ∈ , of a hydraulic actuator. As can be
seen in the oil pressure time series, there are very oscillative behaviors caused by
mechanical resonances. These data have been used in benchmarking studies on
nonlinear system identification [1]. The SOM, MLP and RBF networks are used
as nonlinear identification models to approximate the forward nonlinear dynam-
ics of the hydraulic actuator. These three neural networks are also compared
with the usual linear model, known as the Autoregressive model with Exogenous
Inputs (ARX):

ny −1 n
u −1

ŷ(t + 1) = ai y(t − i) + bj u(t − j) (9)
i=0 j=0

where ai and bj are the coefficients of the model and ŷ(t + 1) is the estimated
value for the plant output at time step t + 1. The coefficients are computed by
978 G. de A. Barreto and A.F.R. Araújo

the Least-Squares Method [6]. The approximation accuracy is evaluated through

the Root Mean Square Error (RMSE):

 M −1
 1 
RM SE =  (y(t + 1) − ŷ(t + 1))2 (10)
M t=0

where M is the length of the estimated series. The data are presented to the four
models without any preprocessing stage. A total number of N = 1024 samples
are available for both the input and output variables. The first 512 samples are
used to train the three networks and to compute the coefficients of the linear
ARX model, while the remaining 512 samples are used to validate the four
models. A training epoch is defined as one presentation of the training samples.
For all the simulations, it is assumed ny = 3 and nu = 2, as suggested in [1].

1.5 4

3
1

0.5
1
valve opening

oil pressure

0 0

−1
−0.5

−2

−1
−3

−1.5 −4
0 100 200 300 400 500 600 700 800 900 1000 0 100 200 300 400 500 600 700 800 900 1000
time time

Fig. 1. Measured values of valve position (left) and oil pressure (right).

The SOM network has six input units, since dim(xin )+dim(xout ) = 5+1 = 6,
and an output layer with 500 neurons. The weights are randomly initialized
between 0 and 1, and adjusted for 600 epochs. The training parameters are the
following: η0 = 1.0, ηT = 10−5 , σi = 250, σT = 10−3 and T = 600×512 = 3×105 .
The time series generated by the ARX and the SOM model during validation
are shown in Figure 2. In these Figures one can note that the estimated values
provided by the linear ARX model were not very accurate (RM SE = 1.0133),
while the accuracy of the SOM was much better (RM SE = 0.2051).
The MLP network has five input units, since dim(xin ) = 5, one hidden
layer with ten neurons and one output neuron. The neurons in the hidden layer
have hyperbolic tangent transfer functions, while the output neuron has a linear
transfer function. The MLP network is trained with backpropagation algorithm
with momentum. The values for the learning rate and the momentum factor are
set to 0.2 and 0.9, respectively. The training is stopped if RMSE ≤ 0.001 or
a maximum number of 600 training epochs is reached. The RBF network also
has five input units, an intermediate layer with neurons with Gaussian basis
 Nonlinear Modeling of Dynamic Systems with the Self-Organizing Map 979

4 4

3 3

2 2

1 1
oil pressure

oil pressure
0 0

−1 −1

Observed data Observed data

−2 Estimated data −2 Estimated data

−3 −3

−4 −4
0 50 100 150 200 250 300 350 400 450 500 0 50 100 150 200 250 300 350 400 450 500
time step time step

Fig. 2. Simulation of the ARX (left) and the SOM (right) models on validation data.
Solid line: estimated signal. Dotted line: observed (true) data.

function, and one output neuron. Following the RBF design in [7], the number of
neurons in the intermediate layer is the same as the number of training samples,
and hence there is a Gaussian kernel centered at every training vector. The
intermediate-to-output weights are just the target values, so the output is simply
a weighted average of the target values of training cases close to the given input
case. The only weights that need to be learned are the radii of the Gaussian
kernels. However, in this paper, this parameter was deterministically varied to
evaluate its effect in the accuracy of the approximation. The statistics describing
the results produced by the three networks are shown in Table 1. For the SOM
and the MLP were performed ten training runs, and for the RBF network the
radius of the Gaussian kernels was varied from 0.1 (minimum RMSE) to 1.0
(maximum RMSE) in increments of 0.1.

Table 1. RMSE values for the SOM, MLP and RBF networks in the identification of
the forward dynamics of the hydraulic actuator.

Forward Modeling
√
RMSE Min Max Mean var
MLP 0.1162 0.2493 0.1554 0.0457
SOM 0.2051 0.2665 0.2259 0.0215
RBF 0.2067 0.4103 0.2994 0.0774

In this table, one can note that the MLP network provides the best results
in general. The SOM algorithm, in its turn, produces better results than a RBF
network with approximately the same number of neurons. An interesting result
is that the SOM is less sensitive to weight initialization than the MLP network,
as can be seen in the fifth column of Table 1. This sensitivity is measured through
980 G. de A. Barreto and A.F.R. Araújo

the standard deviation of the RMSE values generated for the 10 training runs.
A particular difficulty found in designing the MLP network is the occurrence of
overfitting during training and validation. This characteristic is strongly related
to the choice of the number of hidden neurons, and can result in poor perfor-
mance (high RMSE) during model validation. The SOM network does not suffer
from this drawback, since it is in essence a type of vector quantization algorithm:
the more the SOM network is trained, the more precise is the approximation (in
a statistical sense) of the probability distribution of the training data. However,
after some training time, learning in the SOM stabilizes around RMSE=0.20
and no substantial reduction in RMSE is noted.

4 Conclusion

The simulations shown in this paper illustrate the potential of the VQTAM
technique. Additional tests should be performed, such as residual analysis of the
estimation error, noise and fault tolerance, etc., to demonstrate effectively the
viability of using the SOM algorithm in identification and control of nonlinear
dynamic systems. Currently, research is being conducted with the aim of design-
ing a predictive nonlinear controller using the SOM and the VQTAM approach.
Also, a theoretical analysis intends to show that the SOM algorithm, using the
VQTAM scheme, can be used for universal function approximation.

Acknowledgements. The authors thank FAPESP for the financial support

(grant Nr. 98/12699-7).

References
1. Sjöberg, J., Zhang, Q., Ljung, L., Benveniste, A., Deylon, B., Glorennec, P.-Y., Hjal-
marsson, H., Juditsky, A.: Nonlinear Black-Box Modeling in System Identification:
A Unified Overview. Automatica 31:12 (1995) 1691–1724
2. Kohonen, T.: Self-Organizing Maps. 2nd. edn. Springer-Verlag, Berlin Heidelberg
New York (1997)
3. Walter, J., Ritter, H.: Rapid Learning with Parametrized Self-Organizing Maps.
Neurocomputing 12 (1996) 131–153
4. Norgaard, M., Ravn, O., Poulsen, N. K., Hansen L. K.: Neural Networks for Mod-
elling and Control of Dynamic Systems. Springer-Verlag, Berlin Heidelberg New
York (2000)
5. Araújo, A. F. R., Barreto, G. A.: Context in Temporal Sequence Processing: A Self-
Organizing Approach and Its Application to Robotics. IEEE Trans. Neural Nets.
13:1 (2002) 45–57
6. Ljung, L., Glad, T.: Modeling of Dynamic Systems. Prentice-Hall, Englewood Cliffs
NJ (1994)
7. Specht, D. F.: A Generalized Regression Neural Network. IEEE Trans. Neural Nets.
2:5 (1991) 568–576