Regularized Adaptation of Fuzzy Inference Systems. Modelling the Opinion of a Medical Expert about Physical Fitness: An Application MOHIT KUMAR [email protected]Institute of Occupational and Social Medicine, Faculty of Medicine, University of Rostock, D-18055 Rostock, Germany REGINA STOLL [email protected]Institute of Occupational and Social Medicine, Faculty of Medicine, University of Rostock, D-18055 Rostock, Germany NORBERT STOLL [email protected]Institute of Automation, Department of Electrical Engineering and Information Technology, University of Rostock, D-18119 Rostock-Warnemnde, Germany Abstract. This study presents a new approach to adaptation of Sugeno type fuzzy inference systems using regularization, since regularization improves the robustness of standard parameter estimation algorithms leading to stable fuzzy approximation. The proposed method can be used for modelling, identification and control of physical processes. A recursive method for on-line identification of fuzzy parameters employing Tikhonov regularization is suggested. The power of approach was shown by applying it to the modelling, identification, and adaptive control problems of dynamic processes. The proposed approach was used for modelling of human-decisions (experience) with a fuzzy inference system and for the fuzzy approximation of physical fitness with real world medical data. Keywords: fuzzy approximation, model identification, ill-posed, regularization, nonlinear least squares, inverse control, generalized predictive control 1. Introduction Fuzzy systems have been widely applied for modelling complex processes and decision problems (Zadeh (1973)). The main advantage of fuzzy inference system over classical learning systems and neural networks is its linguistic interpretability of its function. Automatic construction or tuning of the fuzzy systems from example data has been widely explored for Sugeno type fuzzy inference system by Babuska (2000), Bodenhofer and Bauer (2000), Espinosa and Vandewalle (2000), and Setnes et al (1998). The Sugeno type fuzzy systems are supposed to ideally combine simplicity with good analytical properties (Takagi and Sugeno (1985)). The general problem of estimating the parameters describing Sugeno fuzzy system from data is ill-posed (Burger et al (2002)). Therefore, nonlinear regularization theory has been employed and a numerical method based on generalized Gauss-Newton was suggested for solving nonlinear regularized optimization problem by Burger et al (2002). In this study, a recursive method for online learning of fuzzy system employing Tikhonov regularization is suggested. This method is based on recursive solution of a nonlinear least squares problem. The antecedent parameters are estimated by Fuzzy Optimization and Decision Making, 2, 317 – 336, 2003 # 2003 Kluwer Academic Publishers. Printed in The Netherlands.
20
Embed
Regularized Adaptation of Fuzzy Inference Systems. Modelling the Opinion of a Medical Expert about Physical Fitness: An Application
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
Regularized Adaptation of Fuzzy Inference Systems.Modelling the Opinion of a Medical Expert aboutPhysical Fitness: An Application
The suggested algorithm was used for on-line tuning of a fuzzy inference system having
49 fitting parameters (32 knots and 17 consequent parameters) with the value of � ¼ 0.001,
�1 ¼ 0.001, and � ¼ 104. Figure 1 shows the initial shape of membership functions. The
input was a sinusoidal signal u(k) ¼ cos(2pk/50). The adaptation of the system was
stopped at k ¼ 25. For the purpose of testing, u(k) was changed to sin(2pk/250) at k ¼ 25
and was further changed to 0.5 sin(2pk/250) þ 0.5 sin(2pk/25) at k ¼ 250. Figure 2 shows
that fuzzy inference system follows the plant function f (u) even after 25th step when
adaptation was stopped. During adaptation, the interpretability of fuzzy system remains
preserved as a result of solving constrained optimization problem at each step.
In Jang et al (1997), the unknown function was identified on-line using ANFIS. For the
on-line learning of ANFIS consequent parameters, recursive least squares estimation
KUMAR, STOLL AND STOLL322
(LSE) with forgetting factor equal to 0.99 was employed and antecedent parameters were
adapted with a learning rate of 0.1 using steepest descent. Table 1 summarizes the
comparison with multilayer perceptron (MLP) and ANFIS. The first two rows of Table 1
are from Jang et al (1997).
The proposed method uses minimum data set (one tenth of ANFIS) for approximating
the unknown function, however with higher number of parameters than ANFIS. The
higher number of parameters can be justified by considering the interpretability of fuzzy
inference system. The fuzzy rules generated by ANFIS are very difficult to interpret
because of the form of consequents (linear combination of inputs) but the rule base
generated by proposed method is quite interpretable, since consequents are real numbers.
Also, the proposed method provides robust and stable fuzzy approximation with uncertain
identification data.
4.2. Predicting Chaotic Time Series
To demonstrate the fast learning and good generalization capability of this algorithm,
again an example already considered by Jang et al (1997) has been studied. It is shown that
Figure 1. Initial guess about shape of membership functions in identifying function f(u).
REGULARIZED ADAPTATION OF FUZZY INFERENCE SYSTEMS 323
trained fuzzy inference system can be used to predict the future values of a chaotic time
series. The performance obtained in this example is compared with other methods. The
time series is generated by simulating the chaotic Mackey-Glass differential delay equation
i.e.
dx
dt¼ 0:2xðt � 17Þ
1þ x10ðt � 17Þ � 0:1xðtÞ;
with xð0Þ ¼ 1:2; xðtÞ ¼ 0 ; for t < 0:
The aim of the problem is to predict the value of x(t þ 6) by using a set of past values i.e.
[x(t � 18), x(t � 12), x(t � 6), x(t)]. Fourth-order Runge-Kutta method was used for the
simulation of above equation and 250 data pairs were extracted for the training set of fuzzy
inference system. Checking data set for validating the identified fuzzy system consists of
500 data pairs which is same as in Jang et al (1997), hence allowing the direct comparison
with other techniques. The number of membership functions assigned to each of four
inputs was set to three. Figure 3 depicts the initial membership functions for each input
Figure 2. On-line identification of function f(u).
KUMAR, STOLL AND STOLL324
variable. Such a fuzzy system was tuned on-line till 250 steps having 97 fitting
parameters (16 knots and 81 consequent parameters) with � ¼ 0.5, �1 ¼ 0.001, � ¼104, and afterwards it was used to predict 500 points of checking data set. To measure the
prediction error non-dimensional error index (NDEI) is defined as the root mean square
divided by the standard deviation of the target series. Figure 4 shows the generalization
capability of regularized fuzzy system learned on-line. Results shown in Table 2 (last four
rows are from Jang et al. (1997)) indicate that regularized learning of fuzzy systems
requires minimum training data for the fuzzy approximation. We find that the prediction
error of ANFIS is smaller than the error of regularized adaptation method. However,
Table 1. Comparison with other identifiers.
Method Parameters number
Time steps
of Adaptation
Backpropagation MLP 261 50000
ANFIS 35 250
Regularized Fuzzy Adaptation 49 25
Figure 3. Initial guess about shape of membership functions for all four inputs in predicting time series.
REGULARIZED ADAPTATION OF FUZZY INFERENCE SYSTEMS 325
we should note that regularized adaptation is on-line method while ANFIS was trained
off-line with 499.5 epochs and using double amount of identification data. Moreover,
regularized adaptation method uses smaller number of parameters and is more interpret-
able than ANFIS.
4.3. Adaptive Control
The feature that the solution of a regularized optimization problem depends on data in a
continuous and stable way, can be exploited for the adaptive control of a time-varying
plant. The regularized learning of fuzzy inference system enables it to learn the changes
occurring in the plant smoothly. Suppose that we have a data set consisting of m
Table 2. Comparison with other predictors.
Method Training cases Parameters number NDEI
Regularized learning(on-line) 250 97 0.065
ANFIS 500 104 0.007
AR model 500 45 0.19
Cascaded-correlation NN 500 693 0.06
Backpropagation MLP 500 540 0.02
Figure 4. Prediction made by regularized learned fuzzy inference system.
KUMAR, STOLL AND STOLL326
observations [(xi, yi)]i¼1,2, . . . ,m which serves as training data for the learning of fuzzy
system. The problem consist of minimizing the functional
Assume that the function Bj satisfy the Lipschitz-estimate
jBjðu; tÞ � Bjðbu; tÞj < �Lju� buj;for all u, ub and t with some nonnegative real constant L. Let (�k, t k ) be the solution of
minimization problem (3) with training data [Xk, Yk], resulting the fuzzy approximation
of desired function (model of plant or controller) at any time. Any model based stable
control strategy can be used to control the plant at this time. Now, suppose that the
dynamics of plant changes and we collect a new data set [X, Y ], consisting of m
observations. Again, we solve the minimization problem (3) with new training data set
[X, Y ]. The stability and convergence analysis of this problem states following results
from Burger et al (2000).
Preposition (Stability) 1. Let �1 > 0, Yk ! Y and Xk ! X. Then the according sequence
of minimizers (�k, t k ) of (3) has a convergent subsequence and the limit of every
convergent subsequence is a minimizer of (3).
The proof of above preposition follows from Burger et al (2000). Let us assume that the
new training data set and old data set satisfy following equations
jjX k � X jj < � k and jjYk � Y jj < � k : ð4Þ
THEOREM (CONVERGENCE) 1. Assume that minimizer of problem (3) exists. Moreover,
let (�k, �k) be a sequence converging to (0, 0) and denote by (�k, t k) the according
sequence of minimizers of (3) with data [Xk, Yk], satisfying (4). Then (�k, t k) has a
convergent subsequence and the limit of every convergent subsequence is a minimizer of
(3) with data [X, Y] if the regularization parameters satisfy
�k1 ! 0; �k
2 ! 0;maxð�k ; �kÞ
�k1
! 0;
9� > 0 : ð�k1=�
k2Þ > �:
Again the proof of above theorem follows from Burger et al (2000). So we note that by
solving a regularized minimization problem, it is possible to get a stable approximation of
an unknown function. Thus the recursive solution of minimization problem (3) allows us
to approximate any time-varying function in a natural stable way, since the solution of
minimization problem depends upon data samples in a continuous and stable manner.
REGULARIZED ADAPTATION OF FUZZY INFERENCE SYSTEMS 327
The proposed algorithm can be applied as such with a suitable choice of regularization
parameters �1 ¼ �2(k) ¼ � > 0. The regularization parameter �2 is chosen independent of
time, contrary to on-line identification, where regularization parameter increases linearly
with time. The optimal value of regularization parameter is always a compromise between
stability and accuracy. A higher value of regularization parameter provides more damping
to oscillations but may results in larger error. To demonstrate the feasibility of our
approach, we consider a time-varying plant described by following discrete dynamical
equation
yðk þ 1Þ ¼ yðkÞuðkÞ1þ rðkÞy2ðkÞ � tan uðkÞ; ð5Þ
where y(k) and u(k) are the state and control action respectively at time step k and r (k) is a
parameter of plant that changes with time and the variation of this parameter with time is
not known. Among the various choices of control strategies, we consider only Inverse
Control and Generalized Predictive Control.
4.3.1. Inverse Control
The idea behind inverse control is to capture the inverse dynamics of plant using a fuzzy
inference system. The fuzzy model is trained to map the inverse of the function describing
the dynamics of plant to be controlled. The trained fuzzy system is then used to generate
control actions for the given desired output. The learning of a fuzzy system can continue
with time, enabling it to be used in adaptive context, when the plant is time varying. Let us
consider the plant described by equation (5) and goal is to track the reference trajectory
given by equation
yrðk þ 1Þ ¼ 0:6sinðk=11Þ þ 0:2sinðk=3Þ:
Let r(k) varies with time in following manner:
rðkÞ ¼
1 when 0 < k � 50;2 when 50 < k � 100;3 when 100 < k � 150;4 when 150 < k � 200:
8>><>>:The proposed algorithm was used for approximating the time-varying function u(k) ¼fk( y(k), y(k þ 1)) with a fuzzy model. Figure 5 depicts the simulation of adaptive inverse
control strategy. The plant is able to track the desired signal very efficiently even when
dynamics of plant changes at step 50, 100 and 150. Figure 6 shows the error curve.
4.3.2. Generalized Predictive Control
Generalized predictive control (GPC) is known to control non-minimum phase plants,
open-loop unstable plants and plants with variable or unknown dead time. GPC had been
KUMAR, STOLL AND STOLL328
originally developed with linear plant predictor models which lead to formulation that can
be solved analytically. If a nonlinear model is used, a nonlinear optimization algorithm is
necessary. This affects the computational efficiency and performance of the method. For
nonlinear plants, the ability of the GPC to make accurate predictions can be enhanced if a
fuzzy inference system is used to learn the dynamics of plant instead of standard nonlinear
modelling techniques. For time-varying plants, the learning of fuzzy inference system can
Figure 5. Adaptive inverse control.
Figure 6. Error curve ( yr � y).
REGULARIZED ADAPTATION OF FUZZY INFERENCE SYSTEMS 329
continue with time to capture the varying dynamics of plant at least around operating
point. Figure 7 shows the block diagram for fuzzy based GPC. Adaptive GPC consist of
following steps
1. Generate a reference trajectory ( yr). If the future trajectory of yr is unknown, keep yrconstant for future trajectory.
2. Keep the switch S in open position, and predict the future performance ypred of plant
from previously calculated input using fuzzy model of the plant.
3. Calculate a new control input (u*) that minimize the cost function.
4. Close the switch S, resulting in application of optimal input (u*) to plant and also the
adaptation of fuzzy inference system.
The selection of minimization algorithm affects the computational efficiency of the
scheme. The objective function to be minimized at every k th step may be taken as
JðuÞ ¼XN2
i¼N1
½yrðk þ iÞ � ypredðk þ iÞ�2 þ XNu
i¼1
uðk þ iÞ2; Cu � h;
where N1 is minimum costing horizon, N2 is maximum costing horizon, and Nu is control
horizon. The constraints on input are in the form of inequality Cu � h. The problem of
finding optimal u is a non-linear constrained optimization problem.
We consider the problem of controlling the plant described by equation (5) based on
GPC. Let the dynamics of the plant changes with time as follows:
rðkÞ ¼
1 when 0 < k � 40;2 when 40 < k � 80;3 when 80 < k � 120;4 when 120 < k � 160:
8>><>>:
Figure 7. Block diagram for fuzzy based GPC.
KUMAR, STOLL AND STOLL330
Let N1 ¼ 1, N2 ¼ 3, Nu ¼ 3, ¼ 0.001, and constraints on input u(k) are in the form
�0.1 < [u(k) � u(k � 1)] < 0.1. To minimize the cost function J, Gauss-Newton method
was used. Figure 8 shows the set point control of the plant using above strategy. It is clear
from the graph that fuzzy inference system adapted very well to the changing dynamics at
time step 40, 80 and 120.
We state the various advantages of the proposed method, compared with ANFIS with
following comments.
1. The regularized adaptation requires a comparatively smaller set of identification data
for the approximation of any unknown function.
2. The proposed method preserves the interpretability of fuzzy system and manually
designed constraints can be put on the shape of membership functions.
3. The use of regularization in learning of a fuzzy inference system leads to robust and
stable fuzzy approximation with uncertain identification data.
4. The proposed method is capable of handling time-varying situations (Adaptive Control).
The disadvantages of proposed method, compared with ANFIS may include following
comments.
1. In some cases, number of adjustable parameters used by ANFIS may be smaller than
that of the proposed method. The higher number of parameters in such situations is
justifiable, since the rule base generated by the proposed method is more interpretable
in comparison to ANFIS.
2. ANFIS sometimes may have smaller approximation error when it is trained off-line.