IAENG International Journal of Computer Science, 37:1, IJCS_37_1_03

______________________________________________________________________________________

A Matlab/Simulink Toolbox for Inversion of

Local Linear Model Trees
Mirko Nentwig

AbstractModels in the form of characteristic diagrams or more specifically, in the form of engine
operating maps are mostly used in the automobile
industry. This yields a large amount of measurements and involves the use of advanced instrumentations. This paper shows a developed software environment, namely a toolbox for the program Matlab/Simulink developed by company Mathworks. The
name of the toolbox is Inversion of the Local Linear
Model Trees and it basically consists of a local inversion of the Local Linear Model Trees (LOLIMOT).
The importance of the inversion in control problems
is widely known. Neural networks are a very effective and popular tools mostly used for modeling. The
inversion of a neural network produces real possibilities to involve the networks in the control problem
schemes. The developed toolbox is explained with
the help of diagrams and GUI structure from Matlab
which tend to clarify the idea of the program and its
structure. The presentation is organized as a short
tutorial of the toolbox, so that a potential user can
directly understand how to access it. Nevertheless,
formal mathematical equations, concerning the neural networks and membership functions, need to be
explained together with the LOLIMOT structure. To
validate and to clarify the explained toolbox, an example from a system used in the automobile industry
is briefly shown.

Paolo Mercorelli
inversion using a physical model approach is presented.
The model which is adopted suffers from some hard reductions in order to be applied. Nowadays neural networks are often adopted for modeling of complex systems. In the neural network area the inversion problem
attracted a lot of researchers and mathematicians. The
inversion problem is a very hard problem which involves,
in the case of neural networks, the inversion of the membership functions which are non linear functions. In Fig.
1 a schematic structure of a possible control system is
represented. From the picture the importance of the inversion of the neural network in particular for tracking
problems emerges. Due to the inexact description of the
model, its imperfect inversion as well as the presence of
not modeled external disturbances imply that the controller must consist of a structured feedback part. This

Keywords: Neuro-fuzzy identification, nonlinear systems, local linear model trees (LOLIMOT), inversion,
Matlab/Simulink

Figure 1: Complete Control Scheme

Introduction and Motivation

The inversion problem has always been an attractive

problem for the control area. In fact, through an inversion of the model it is possible to do a feed-forward
action which can represent an important part, very often
adopted in modern control techniques. The feed-forward
action is applied, for instance, in tracking problems, stationary correction of the error, control with two degrees
of freedom, estimation of the inputs and in general when
some pre-actions are needed in order to prepare the system with regard to energy. In [5] a controller based on
University of Applied Sciences Ostfalia, Department of Automotive Engineering, Robert-Koch-Platz 10 - 14, 38440 Wolfsburg,
Germany, Tel. +49-(0)5361-831615 Fax. +49-(0)5361-831602, EMail: [email protected], [email protected]

procedure applies the LOLIMOT algorithm [6], which is

based upon Neural-Fuzzy models of Takigi Sugeno type.
During the execution of this algorithm, a divide and
conquer strategy is applied to the modeling problem, so
that the major problem is split into smaller ones.

1.1

Structure

The basic network structure of a local linear neuro fuzzy

model is depicted in Fig. 2. Every neuron consists of a
local linear model, a so called LLM, and a validity function , which defines the validity of the LLM within the
input space.
The local linear model output is defined by:
yi = wi0 + wi1 u1 + wi2 u2 + ... + wip up

(Advance online publication: 1 February 2010)

IAENG International Journal of Computer Science, 37:1, IJCS_37_1_03

______________________________________________________________________________________

Figure 2: Left: Network Structure of Local Linear Neuro Fuzzy Model. Right: Partitioning of Input Space by
Validity Functions

with wij as parameters of neuron i.

If the validity functions are chosen as normalized Gaussians, then:
PM

i=1 i (u) = 1 and i =

(u)
PM i
,
j=1 j (u)

where the membership function is defined as




1 (u1 ci1 )2
1 (u2 ci2 )2
i (u) = exp
+exp
2
2
2
i1
2
i2
!
1 (up cip )2
+ ... + exp
(1)
2
2
ip


1.2

Estimation of Linear Parameters

Xi =

1
1
..
.

u1 (1)
u1 (1)
..
.

...
...

up (1)
up (1)
..
.

1 u1 (N ) u2 (N ) . . .

up (N )

wi0
wi1

wi = .
..
wip

and

It should be noted that the regression matrices of all linear models i = 1, ..., M are identical, because all the elements of X i do not depend on i. Due to the structure
of the local linear models with their limited validity, its
useful to apply a weighted least squares. The validity factors are described by the validity functions. This leads
to the following N xN weight matrix

i (u(1))
0
...
0

0
i (u(1)) . . .
0

Qi =
.
..
..
.
..
..

.
.
.
0

The estimation of the linear equation parameters is done

through an optimization by local least squares method.
This method offers a fast possibility for estimating the
rule conclusions. The idea of the local estimation is to
treat the optimization problem of every conclusion as an
individual one. Instead of estimating all n = M (p + 1)
parameters parallel, M single local estimations for every
p + 1 parameters of the model are carried out.
The parameter vector for each i = 1, ..., M , and the
regression matrix are respectively:

u2 (1)
u2 (1)
..
.

...

i (u(N ))

The weighted least squares solution of the rule conclusion

parameters is given by


w
i = X Ti Qi X i X Ti Qi y.
This estimation has to be carried out for all i = 1, ..., M
models.

1.3

Structure
Premises

Optimization

of

Rule

The local least squares method mentioned for estimating

the rule conclusions is only applicable if the validity functions have been estimated first.
The number of the validity functions and their parameters define the partitioning of the input space. The validity functions are depicted in Fig. 2. These divide the

(Advance online publication: 1 February 2010)

IAENG International Journal of Computer Science, 37:1, IJCS_37_1_03

______________________________________________________________________________________

input space in rectangle areas shown in Fig. 2. If normalized Gaussians are used, then cij describe the centers
of the rectangles and the standard deviations ij describe
the extension in each dimension. The proportional factor
k relates the standard deviation of the validity functions
to the extension of the rectangle by ij = k ij .

1.4

carried out. By each division the following steps are

carried out
3a: Construction of the n-dimensional membership functions for every hyper rectangle
3b: Construction of the validity functions
3c: Local estimation of the parameters which
belong to the newly created local linear models

Local Linear Model Tree Algorithm

(LOLIMOT)

3d: Calculation of the loss functions for the

complete model

The LOLIMOT is an incremental tree construction

algorithm, which divides the input space axes in an
orthogonal way. By iterations, one new local linear
model is added to the model. To do that, validity
functions are calculated and the local linear models are
adapted with the least squares method. The only fiddle
parameter is k . This is a proportional factor between
the size of the rectangles and the standard deviation.
The optimal value depends on the application, but
generally the following delivers good results: k = 31 .
The Standard deviations are calculated as follows:
ij = k ij , where ij describes the extension of the
hyper rectangles of model I in dimension uj depicted
in Fig. 2. The LOLIMOT algorithm consists of an
external loop (external algorithm) in which the structure
of the model is determined and an inner loop (internal
algorithm) in which the parameters of the model are
estimated by local least squares.
Short Description of the Algorithm
1. Start with an initial model
Construct the validity functions for the beginning
input space portioning and estimate the LLM
parameters by local least squares. Set M to starting
number of LLM. If no starting input space division
is available set M = 1 and start with one LLM,
which covers the whole input space.
2. Find the worst LLM
Calculate the loss function for each of the I =
1, ..., M local linear models. A local loss function
is used, with a squared model error:
Ii =

N
X

e2 (j)i (u(j)).

j=1

Find the worst LLM, where maxi (Ii ). i defines the

index of the worst LLM.
3. Check all possible divisions
The worst LLM i will be considered for further optimizations. Each hyper rectangle of the local linear
models will be divided orthogonally along the axes
in two pieces. Divisions in every of p dimensions are

4. Find the best division

The best among p alternatives of step 3 is chosen.
The validity functions which were constructed in
step 3a and the optimized local linear models in
step 3c, will be added to the model. The number of
LLM are incremented M M + 1.
5. Check for convergence
If the final criteria is met, then stop the algorithm,
otherwise return to step 2
More than one possibility exists for the final criteria,
e.g. maximal model complexity that means maximal
model count, statistical judges or information reliability.
In Fig. 3 the LOLIMOT algorithm for the first five iteration steps with a two dimensional input space is depicted.
Especially two features make the algorithm extremely
fast. First, during every iteration not all local linear models are considered for the division. Step 2 selects only
the worst local linear model which may gain the biggest
performance improvement. For example in iteration step
four only the LLM 4-4 is considered for further optimization. All other LLMs remain unchanged.
Secondly, in step 3c the local least squares method makes
it possible that only the parameter of the two newly created LLM needs to be estimated. For Example, if in the
fourth iteration step the LLM 4-4 is divided in LLM 5-4
and 5-5 the local linear models 4-1, 4-2 and 4-3 can be
directly taken to the next iteration loop 5-1, 5-2 and 5-3
- without any changes.

LOLIMOT Inversion

In this paragraph the inversion of the local linear model

tree is described. The idea is to develop an algorithm
which returns the required model input ur depending on
the desired model output y and the other model inputs
u. Fink and Toepfer in [1] offer some strategies in their
analysis of the inversion of nonlinear models:
Inverse access by numerical inversion
Only one model of the nonlinear function is created.
This model is used for standard and inverse access.

(Advance online publication: 1 February 2010)

IAENG International Journal of Computer Science, 37:1, IJCS_37_1_03

______________________________________________________________________________________

2.1

The Validity Functions Issue

Consider the model output function y, with M local linear model u = [u1 , ..., up ] inputs
y =

M
X

(wi0 + wi1 u1 + +wi2 u2 + ... + wip up )i (u), (2)

i=1

the validity function

i (u)
i = PM
i=1 j (u)
and the membership function



1 (u1 ci1 )2
+
i (u) = exp
2
2
i1



1 (u2 ci2 )2
exp
+ ...
2
2
i2
1
+ exp
2

Figure 3: LOLIMOT Algorihm

The inverse access equals a numerical inversion and
requires the application of optimization methods to
find the required input for the requested output.
Data driven generation of an inverse model
As well as the model for feed forward access, a model
for inverse access will be created.
Analytical inversion of models
This method applies a direct inversion of the forward
trained model. Hence its an advantage to use model
architectures which allow the direct calculation of
the inverse model by its own parameters.
The developed algorithm applies an analytical/numerical
inversion of a given local linear model structure.
To achieve this goal, the following constraints are needed:

1. An existing forward trained local linear model tree

of the process is available
2. An expected model output y is known
3. The input values of the other inputs on which y is
depending are known

(up cip )2
2
ip

!!
. (3)

A difficulty is that, due to the exponential quadratic nonlinearity, the model is not invertible. Hence, it is necessary to convert the functions into a linear type, as shown
in Fig. 4. The membership function is split into a spline
function that consists of the linear functions.
(
k
1.6r
1.6 r (ur c) + 1 k + c ur c
ir =
k
r
1.6 r (ur + c) + 1 c ur 1.6
k + c.
(4)
Using the function defined in equation (4), it is now possible to invert the local linear model tree.

2.2

Algorithm for the Inversion of the LLM

We apply the following algorithm to invert the model:

1. Calculate the LLMs represented in equation (2),
omitting the required input variable ur . That is, calculate the LLMs with the available input data until
there only remains a linear equation depending on
one input, e.g., yi = wr ur + ucalc . To be more comprehensible, if yi = wi0 + wi1 ui1 + wi2 ui2 + wi3 ui3
and ui1 is required, then yi = wi1 ui1 + ucalc , with
wr ur = wi1 ui1 and ucalc = wi0 + wi2 ui2 + wi3 ui3 .
2. Calculate the membership functions represented in
equation (3), omitting the required input variable ur .
The membership function of the LLMs is calculated
with the available input data to the extent possible.
Since the nonlinear term depending on ur is omitted,
the membership function is a constant number. To
be more precise,



1 (u2 ci2 )2
c (u) = exp
+ ...
2
2
i2
!!
1 (up cip )2
+ exp
. (5)
2
2
ip

(Advance online publication: 1 February 2010)

IAENG International Journal of Computer Science, 37:1, IJCS_37_1_03

______________________________________________________________________________________

as the following:

(ur )i,r =

(ur )i,r,1
(ur )i,r,2 .

(7)

5. In the following loop, consider every interval for a

possible solution of ur .

Determine which of i membership functions

(ur )i,r are valid for the currently considered
interval, by checking every spline. (ur )i,r,1 is
valid if
(interval lef t)i,r,1 0
(interval right)i,r,1 1

(8)

(ur )i,r,2 is valid if

(interval lef t)i,r,2 1
(interval right)i,r,2 0, (9)
where indicates the and logical function.

Figure 4: Top: Linear and nonlinear validity function.

Bottom: Linear functions on the intervals.

Then, the input ur is reconsidered in the final membership function and the following expression is obtained,



1 (ur cir )2
i (u) = exp
)
+ c .
2
2
ir

(6)

3. Create the linear membership function for the

required input as from equation (4).

4. Partition the input space ur .

The input space of ur is partitioned into q search intervals, which are used in the later estimation of ur .
Every interval describes the validity of half of the local linear model. Thus, the input space of every LLM
is divided into two intervals. This is necessary due
to the structure of the new linear membership functions, because they consist of two linear functions as
mentioned above. For every interval, a left function and a right function are considered (Fig. 3).
For the sake of brevity, equation (4) is represented

Use the valid membership spline functions to

create validity functions for each local linear
model. Taking the previously calculated part of
the membership function i, i,calc as in (6), and
sum it with the valid linear spline membership
function i (ur )i,r,{1,2} , where (ur )i,r,{1,2}
represents a valid spline membership function
within the range of functions. This yields:
(ur )i = (ur )i,r,{1,2} + i,calc and
r )i
(ur )i = PM(u(u
.
)
j=1

r j

If there are no valid linear membership functions for a local linear model, then the model
will not be considered for further actions.
Initially create the output function for every local linear model by multiplying its validity function with the local linear model function:
y(ur )LLM,i = y(ur )i (ur )i .
Next, sum the output functions to create the
model output:
y(ur ) =

M
X

y(ur )LLM,i .

i=1

Finally, equate the model output function to

the desired model output value, and solve the
resulting equation with respect to the variable
ur :
y = y(ur ).

(Advance online publication: 1 February 2010)

IAENG International Journal of Computer Science, 37:1, IJCS_37_1_03

______________________________________________________________________________________

Verify that the calculated ur is inside the currently considered interval.

interval lef t ur interval right.
If so, accept it as one possible solution. If not,
disregard it.
Due to the structure of the validity functions, an inversion of the model is possible only within the input space
of the required variable. Beyond these borders, the model
input will drift to zero, which is comparable to the normal LOLIMOT behavior. That is, once a work nominal
point is chosen, the inversion is possible within the input
domain. The worst case is if the working nominal point is
close to the border of the domain. If so, a more suitable
division of the input space is needed.

Toolbox Description

3.1

Setup the Local Linear Model Tree

Structure

y output - Value of the desired model output, e.g.

y output = 123.456
k sigma - The k sigma used at model creation, e.g.
k sigma = 0.33 refer LOLIMOT introduction.
Output Parameters:
possible solutions - Row vector of possible values for the desired input[sol.1 sol.2 ... sol.N ] e.g.
[1, 4, 5].

3.3

An Application

The considered system is a throttle valve shown in

Fig. 6, the parts of the internal combustion engine are
depicted. Particularly difficult is the model of the air
path which consists of an air filter, a throttle valve, a
collecting tank and an intake valve. A complete model
with chemical reaction is not considered. The command

To apply the inversion tool to a LOLIMOT its necessary

to transform the model into the following data structure
as depicted in fig. 5 on page 7. The structures should be
named Model so it is possible to use them also along
with the Simulink block set. The structure Networkname consists of neurons which are named Networkname(1) Networkname(M). Each neuron includes the
parameters for its membership function and the local linear model g. The membership function consists
of the central and standard deviation parameters for every dimension of the input space. The LLM parameter
vector g owns the weight values of the local liner model
[w0 w1 ... wn ].

3.2

Usage of the Matlab Command Line Inversion Tool

The usage of the command line inversion tool is intuitive.

To call the inversion function you have to type inver on
the Matlab command line. Before executing the function
you have to submit some parameters.
Matlab Command:
[possible solutions] =
inver(Lolimot, u seek, u input, y output, k sigma)

Input Parameters:
Lolimot - Local linear model tree structure
u seek - Number of the desired input e.g. 1 = u1
u input - Model input data as row vector
[u1 u2 ... un ], where the desired input should
be set to 0, but it is a not caring value, e.g.
[u1 u2 (required) u3 ], i.e. [12.4 0 34.2]

Figure 6: Top: Overview. Bottom: Schematic Structure

of Throttle Valve

line function will be applied to a manifold air mass flow

model with a size of 3 neurons, which depends on four
inputs, where u1 represents the throttle angle in degrees,
u2 is the ambient air pressure in mbar, u3 represents
the manifold air pressure in mbar and u4 the ambient
temperature in Celsius, see the lower part of Fig. 6. The
model output is manifold air mass flow in kg/h. The

(Advance online publication: 1 February 2010)

IAENG International Journal of Computer Science, 37:1, IJCS_37_1_03

______________________________________________________________________________________

Figure 5: Local Linear Model Structure

model was trained with a k sigma equal to 0.33.
With the following data set, all four required input
values for the requested model output are estimated.
They are depending on the remaining three inputs and
the model output.
u = [40.32 1012.5 977.3 21.21];
y = 147.18.
The known but omitted input value for the required
input will be used for the validation of the estimated
value.
Estimation of Input u1 . First the estimation of
throttle angle u1 is carried out. The input parameters
are: the model structure named Model, the required
input u1 = 1, the measured input data, the desired
model output value and the k sigma which was used for
model training. As Output value, u1 is selected.
Matlab Command:
u1 = inver(M odel, 1, [0 1012.5 977.3 21.21], 147.18, 0.33)
u1 = 39.1160

The estimated throttle angle is 39.116 the reference

angle is 40.32 , which results in a difference of 3.08%.
Estimation of Input u2 .
The u2 input represents the ambient air pressure. The input parameters
are those of the Model again but the required input is
set to u2 = 2 and adapted to the input data. All other
parameters remain the same.
Matlab Command:
u2 = inver(M odel, 2, [40.32 0 977.3 21.21], 147.18, 0.33)
u2 = 1.0e + 003 [1.116 1.0116]

The inversion tool estimated two possible solutions for

the ambient pressure of 1011.6 mbar or 1166 mbar. The
real value is 1015.5 mbar. Due to the difference of the
results 1011.6 mbar is assumed to be the right result. In

this case the difference is 0.39 %.

Estimation of Input u3 .
For the estimation of
the manifold air pressure u3 the required input must be
changed and the input vector is set to [40.32 1012.5
0 21.21] the other values must not be changed.
Matlab Command:
u3 = inver(M odel, 3, [40.32 1012.5 0 21.21], 147.18, 0.33)
u3 = 982.9629

The reference pressure for the requested model output is

977.3 mbar, as a result an air pressure of 982.96 mbar is
obtained. The difference is 0.58%.
Estimation of Input u4 .
At last the ambient
temperature u4 must be estimated. Now the required
input must be the 4th and the parameter vector must be
adapted.
Matlab Command:
u4 = inver(M odel, 4, [40.32 1012.5 977.3 0], 147.18, 0.33)
u4 = 21.2064

The function estimated an ambient temperature of

21.2064 C. The reference temperature is 21.21 C. In this
case the error seems to be caused by the rounding error.
Therefore the inversion tool matches the real input value
with no difference.
As demonstrated, the command line tool offers a
strong possibility to inverse local linear model trees with
a high accuracy and less time effort. If a more suitable
input value is calculated, an automatic selection criterion
for the possible solution is useful. This can be introduced
by taking the previous process input value ur (k 1) and
calculating the difference with the estimated values, if
the process is not an impulsive process. The value with
the smallest difference should be the right one.

(Advance online publication: 1 February 2010)

IAENG International Journal of Computer Science, 37:1, IJCS_37_1_03

______________________________________________________________________________________

3.4

Usage of the Simulink Inversion Block

The developed inversion block (inver) for Matlab

Simulink implements the features of the inverse local linear model trees into a Simulink Model. This is very important for the application in a control loop.
The necessary settings are similar to the command line
tool settings. The Matlab workspace variable of the
LOLIMOT model has to be named Model.
In the picture below, the Simulink block is depicted. It
consists of three inputs: u prev, u input and y desired
and one output u crtl. The u prev input is used for the
k 1 value of the required input which is taken for the
selection of the valid required value, if more than one solution is possible. The multiplexed model input data is
used for u input. The required input should be left open
as at the command line tool in Matlab, hence its a not
caring input. For input y desired should be provided
the desired value for the model/process output. Finally,
after the inversion of the model, block output u crtl is
the required (desired) input value for the desired output
at the output port of the block.
Similar to the command line version the user must specify
in the dialog box depicted in Fig. 7 the required input
u seek which has to be the number of the inputs and
k sigma which ware used at the training of the local linear model tree. In Fig. 8 a possible application of the

The importance of the paper in the control area consists

of the possibility to use the LOLIMOT through its inversion as a forward part of a controller. The inversion
of a throttle valve feed forward model is demonstrated
and very good results are achieved. Future projects involve the testing of the feed forward control in a feedback
control structure.

Figure 8: Simulink Structure of a Feed Forward Control Scheme Based on the Inversion of a Neural Network
Model

References
[1] A. Toepfer A. Fink. Technical report - on the inversion of nonlinear models. 2003.
[2] R. Isermann A. Fink, S. Toepfer. Neuro and neurofuzzy identification for model-based control. IFAC
Workshop on Advanced Fuzzy/Neural Control, Valencia, Spain, Volume Q:111116, 2001.
[3] R. Isermann A. Fink, S. Toepfer. Nonlinear modelbased control with local linear neuro-fuzzy models.
Archive of Applied Mechanics, Volume 72(11-12):911
922, 2003.

Figure 7: Top: Simulink block. Bottom: Dialog Simulink

structure
toolbox is shown. In the scheme a feedback controller can
also be considered. A sinus function is tried to tracked.

Conclusions and Future Work

This paper deals with local inversion of the Local Linear Model Trees. An analytical solution is presented and
implemented in Matlab. Further the theoretical structures of the neural network and also a useful Matlab tool
for any researcher who utilizes LOLIMOT are explained.

[4] A. Fink and O. Nelles. Nonlinear internal model control based on local linear neural networks. IEEE Systems, Man, and Cybernetics, Tucson, USA, 2001.
[5] P. Mercorelli. An optimal minimum phase approximating pd regulator for robust control of a throttle
plate. 45th IEEE Conference on Decision and Control
(CDC2006), San Diego (USA), 13th-15th December,
2006.
[6] O. Nelles. Nonlinear System Identification with Local
Linear Neuro-Fuzzy Models. Shaker Verlag, 1999.
[7] R. Isermann O. Nelles, A. Fink. Local linear model
trees (lolimot) toolbox for nonlinear system identification. 12th IFAC Symposium on System Identification
(SYSID), Santa Barbara, USA, 2000.

(Advance online publication: 1 February 2010)