See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/220795489

Limitations of multivariable controller tuning using genetic algorithms

Conference Paper · June 2009

DOI: 10.1145/1731740.1731790 · Source: DBLP

CITATIONS READS

9 1,391

2 authors, including:

Karl O Jones
Liverpool John Moores University
76 PUBLICATIONS 570 CITATIONS

SEE PROFILE

All content following this page was uploaded by Karl O Jones on 31 May 2014.

The user has requested enhancement of the downloaded file.

 International Conference on Computer Systems and Technologies - CompSysTech’2009

LIMITATIONS OF MULTIVARIABLE CONTROLLER TUNING USING

GENETIC ALGORITHMS

Dr. Karl O. Jones and Wilfried Hengue

Abstract: In recent years Evolutionary Computation has come of age, with Genetic Algorithms (GA)
being possibly the most popular technique. A study is presented revealing the performance of a GA in
determining the PID tuning parameters for a multivariable process, including decoupling controllers. The
process used for this investigation is a distillation column which is a MIMO high-order, nonlinear system. The
results indicate some limitations of using GAs for controller tuning when MIMO systems are involved.
Keywords: Genetic algorithm, multivariable, PID, tuning.

INTRODUCTION
The most common controller used in industry is the Proportional-Integrate-Derivative
(PID) controller. In order for the PID controller to work properly it is essential that its
parameters are tuned, and there are numerous methods that can be used, such as
Zeigler-Nichols [1] or Cohen-Coon [2], both of which are widely used experimental
approaches. While these tuning techniques are wholly suitable for single-input single-
output (SISO) systems, their use becomes extremely problematic when a multi-input multi-
output (MIMO) system is considered. A characteristic feature of MIMO systems is process
interaction where each manipulated variable can affect each output variable. Control of
multi-input multi-output (MIMO) processes can be achieved either by using a set of
independent SISO controllers or by employing a centralized multivariable controller.
Although centralized multivariable control has a number of benefits, multiple SISO
controllers are often employed to control interacting multivariable processes because of
their simplicity. The benefits of multiple SISO controllers over centralized control include
fault tolerance, simplified design, operation flexibility and so on [3]. It should be noted that
because of the single loop structure, a decentralized controller cannot remove process
interactions [4]. The problem on interaction can be reduced with the inclusion of a
decoupling element in the control system, although this further increases the complexity,
for example an nxn process requires a total of n PID controllers and n decouplers.
Simultaneously tuning a number of PID controllers can prove extremely difficult and very
time consuming.
One possible resolution to this manual tuning process is to utilise an automatic
optimisation process. Traditional optimisation approaches can become trapped in local
minima, however newer approaches such as Evolutionary Computation involve large
search populations and thus have more opportunity to find a global optima. Evolutionary
systems have been considered as an optimisation tool since the early 1950s, and until the
1960s, the field of evolutionary systems was working in parallel with research into Genetic
Algorithm (GA). When the two areas began interacting, the field of evolutionary
programming appeared, introducing concepts of evolution, selection and mutation. Holland
[5] defined the concept of the GA as a metaphor of the Darwinian theory of evolution. GAs
view learning in terms of competition amongst an evolving population. GAs have been
applied to numerous problems, and are known as an efficient optimisation method that is
commonly used in industry.
GAs encode potential solutions to a problem on a simple chromosome-like data
structure and apply recombination operators to these structures to preserve critical
information. A GA implementation begins with a population of chromosomes, often
random, which is then evaluated and reproductive opportunities allocated so that those
chromosomes which represent a better solution to the problem are given more chance to
“reproduce”. The suitability of a solution is usually defined with respect to the current

- -
 International Conference on Computer Systems and Technologies - CompSysTech’2009
population [6]. GA techniques have a solid theoretical foundation [5], based on the
Schema Theorem [6].

BINARY DISTILLATION COLUMN

The Binary Distillation Column separates a mixture of two components having
different boiling points, with the process enhanced by making separation occur in stages
within the column. One way to achieve this is by incorporating a liquid stream of high purity
distillate (known as reflux) back into the column through a series of sieve trays. Although
there are benefits to using reflux, the resultant process dynamics creates an extremely
complex control system. The distillation system has two inputs, S and R, and two outputs, B
and D (Figure 1), where S is the hot steam used to heat the column, B are the impurities
being rejected to R which in turn are the impurities being returned to the column in order to
make them pure, finally D is the pure liquid obtained from the column

Figure 1 Distillation Column

The system transfer function has been modelled by Wood and Berry [7] as:
  12.8e  s  18.9e  3s 
 y1(s )   16.7s  1 
21.0s  1   u1(s ) 
   7s   (1)
 y 2 (s )  6.6e  19.4e  3s  u 2 (s )
 
 10.9s  1 14.4s  1 
where y1(s) and y2(s) are the overhead and bottom compositions of methanol, respectively;
u1(s) is the reflux flow rate and u2(s) is the steam flow rate to the reboiler; d(s) is the feed
flow rate, a disturbance variable. The time constants are in minutes. The process can be
represented by the block diagram shown in Figure 2. Controllers D11 and D22 are the
primary control for the system, while D12 and D21 are the decoupling controllers, r1 and r2
are the input references, u1 and u2 are the combined controller signals, and y1 and y2 are
the outputs.

Figure 2 Distillation Column model with PID Controllers and GA for Tuning.

- -
 International Conference on Computer Systems and Technologies - CompSysTech’2009

GENETIC ALGORITHM OPERATION

The Genetic Algorithm (GA) is a powerful and broadly applicable stochastic search
and optimisation technique, and is probably the most widely recognised evolutionary
computation technique. A GA is an iterative process, with an iteration known as a
generation. Common practice is to terminate a GA after a specified number of generations
and then examine the best chromosomes. If no satisfactory solution is determined, then
the GA can be restarted.
Table 1 Genetic Algorithm process
1. Randomly generate an initial population P(0)
2. Fitness evaluation F(i) for each individual i in the current
population P(t)
3. Do a stochastic selection S(i) for each individual i in the population
P(t) in order to have S(i) proportional to F(i)
4. Generate P(t+1) by probabilistically selecting individuals from P(t)
to produce offspring via genetic operators
The process repeats until the maximum number of generations is reached

Importance of the Fitness Function

The fitness function measures the quality of the represented solution and it is always
dependant on the problem under consideration. Seen as the most important parameter
within any optimisation system and in particular in a GA, the fitness function determines
how individuals are selected to provide for the next generation.
As stated the fitness function depends on the problem under consideration, hence
certain cost functions will not provide a good solution for some systems but for others,
which have a different structure. A poorly selected fitness function will not provide an
optimal solution.

APPLICATION RESULTS
The distillation process described by equation (1) was simulated within Simulink in
MATLAB. The GA was executed using a population of 50 over 1000 generations. The GA
was designed to determine 10 parameters which represented the PID parameters for D11
and D22 as well as the PI values for D12 and D21. The fitness function employed was based
on the Integral Absolute Error (IAE) of each process output:
= ( ∫| − |) + ( ∫| − |) (2)
where Sp1 and Sp2 refers to the two set-point signals, and a and b are weighting factors.
While most workers use a step input, this work uses a square wave input profile to ensure
that the developed controller is capable of maintaining regulation for both positive and
negative step changes.
The first tests involved setting all the process time-delays to zero and removing the
decoupling controllers. The resulting transient response (Figure 3) shows reasonable
control for both outputs, and highlights that the GA is capable of finding suitable PID
parameters. Next the GA was applied to the process with the delays included (no
decoupling controllers). Figure 4 illustrates that the developed PID controllers can provide
some form of control, however there is evidence of the time-delays reducing the
effectiveness of the controllers. There is clear interaction between the two loops, however
since there are no decoupling controllers this is only to be expected.
The next test included both the process time-delays and the decoupling controllers
(D12 and D21). In this, instance the GA determined controller parameters that produced an
acceptable controlled performance (Figure 5), limited owing to loop interaction. It can be
stated, that the decoupling controllers are only performing an adequate function. During

- -
 International Conference on Computer Systems and Technologies - CompSysTech’2009
another test the PID parameters for the four controllers produced a response that
deteriorates as time progresses (Figure 6). This would suggest that having considered
some 50,000 sets of PID values during its operation, the GA has not been able to
determine a set that is optimal.

DISCUSSION
Throughout the literature there are numerous cases of genetic algorithms being able
to determine the optimal set of parameters for a given problem. Furthermore, GAs are
frequently recommended as the most appropriate tool for any optimisation problem. The
results presented here suggest that there is a limitation to the capabilities of GAs.
It is not clear where the problem lies: is it a limitation of the genetic algorithm as an
optimisation methodology, or is the failing down to an inadequately designed fitness
function? The latter might seem the most likely problem area, however the fitness function
used in this work was designed to minimise the error between the loop set-points and
outputs and it is clear from the result in Figure 6 that there is a significant level of error
during a steady set-point period.

CONCLUDING REMARKS
This work has focused on how to tune a MIMO process. Classical methods like
Zeigler-Nichols have been used widely within the industry for tuning controllers. The
method is robust and widely used because of its online tuning but it is also described as
old fashioned and an aggressive tuning which might create error. The aim of a control is
not to create errors but to control them.
It is not clear that the application of GAs to tuning a set of PID controllers on a MIMO
system is wholly successful. The results presented for a system with no time-delays is
quite successful, as is the application to a process with decoupling. However, the results
for the complete distillation model (which involves decoupling controllers and time-delays
on the process signals) indicate that the GA has not determined optimal values. The
reasons behind this lack of success need further investigation: firstly, is the full process not
appropriate for PID controllers? Secondly, would an alternative more complicated fitness
function provide improved results?

REFERENCES
[1] Zeigler, J. G. and Nichols, N. B. 1943. Process lags in automatic-control circuits.
Transactions of the ASME, volume 65, p. 433–444.
[2] Cohen, G. H. and Coon, G. A. 1953. Theoretical Consideration of Related Control.
Transactions of the ASME, vol. 75, p. 827-834.
[3] Campo, P. and M. Morari. 1994. Achievable closed-loop properties of systems under
decentralized control: conditions involving the steady-state gain. IEEE Transactions on
Automatic Control. 39 (5), p. 932-943.
[4] Garelli, F., R.J. Mantz and H. De Battista. 2006. Limiting interactions in decentralized
control of MIMO systems. Journal of Process Control. Vol. 16. p. 473-483.
[5] Holland, J. H. 1975. Adaptation in Natural and Artificial Systems. University of
Michigan Press, USA.
[6] Whitley, D. 1993. A Genetic Algorithm Tutorial. Computer Science Department,
Colorado State University Fort Collins, CO 80523, USA.
[7] Wood, R. K. and Berry, M. W. 1973. Terminal Composition Control of a Binary
Distillation Column, Chemical Engineering Science, Vol. 28,p. 1707-1717.

- -
 International Conference on Computer Systems and Technologies - CompSysTech’2009
ABOUT THE AUTHORS
Dr. Karl O. Jones is a Principal Lecturer in the School of Engineering at Liverpool
John Moores University, while Wilfried Hengue was an MSc student working under his
supervision. Having gained his MSc, Wilfried now works in France.
Phone: +44 151 231 2199, Е-mail: [email protected].

RESULT FIGURES

1.2 1.2
Output 1
Output 1
Output 2
1 Output 2
1 Set point 1
Set point 1
Set point 2
0.8 Set point 2
0.8
0.6

0.6 0.4

Output
Output

0.4 0.2

0
0.2
-0.2

0
-0.4

-0.2 -0.6
0 10 20 30 40 50 60 0 5 10 15 20 25 30 35 40
Time (min) Time (min)

Figure 3 Process with no delays and Figure 5 Process with delays and
no decoupling controllers. decoupling controllers included.

1.2 Output 1 3
Output 2 Output 1
Set point 1 2.5 Output 2
1
Set point 2 Set point 1
2 Set point 2
0.8

1.5
0.6
Output
Output

0.4
0.5

0.2
0

0 -0.5

-0.2 -1
0 5 10 15 20 25 30 35 40 0 5 10 15 20 25 30 35 40 45 50
Time (min) Time (min)

Figure 4 Process with no delays Figure 6 Deteriorating controlled

(decoupling controllers included). response (process with delays and
decoupling).

- -

View publication stats