Вы находитесь на странице: 1из 5

Using Evolutionary Techniques to Design Control System For Non-linear Systems 1

Using Evolutionary Techniques to Design


Control System for Non-linear Systems
 Mr Joe Bloggs

 A simple example of a nonlinear system is an inverted


Abstract— Many of the techniques taught in undergraduate pendulum on a free-moving cart [2, 3]. More complex
control systems courses rely on the systems being analysed to be examples can be found in many types of aircraft [4-6], in
linear, or that a linear approximation can be made of the system. industrial furnaces [7], power plants[8] and many other areas.
However, there are many occasions where we want to design a
control system for a non-linear system. In these cases we need
For example, Xioa and Jiang[9] give the following equations
different tools. for a continuous stirred tank reactor:
Evolutionary Techniques (ETs) are a group of algorithms and     1 
heuristics that draw inspiration from genetics and evolutionary
process in biology. They offer a way generating optimised (1)
solutions to non-linear systems. ETs have been used numerous
times to attempt to design control systems. This paper looks at a   1   1   
selection of instances where ETs have been used to design control
systems, what the system was that they were trying to control and (2)
how successful they were in designing a controller for that  
system. (3)
where , , and are constants and  is a function
Index Terms— control systems, evolutionary techniques,
genetic algorithms, non-linear systems defined as:


I. INTRODUCTION 1  
URING undergraduate studies in control systems, content (4)
is often limited to the study of linear systems. In As can be seen, the nonlinearities in these equations would
practice, a control engineer may often find that they need to make the use of standard control engineering techniques
design controller for systems that are nonlinear. In these impossible to use.
cases, there is a need for additional tools to be used. 2) Implications of nonlinearity
Evolutionary techniques offer a powerful toolset for solving The most important impact of nonlinearity is that we cannot
optimisation problems. They can be applied to nonlinear apply standard control system techniques to the analysis and
systems as easily as linear ones and offer a way to explore design of controllers for such a system without devising a
many possible solutions to a problem simultaneously to find linear approximation of the system [1]. A linear
an optimal solution. As such, they are worth evaluating as approximation may not be possible for the full range of
possible methods for developing controllers for nonlinear possible inputs and may contain significant errors.
systems. Having established that there are many non-linear control
systems that we are interested in that we cannot use linear
II. NON LINEAR CONTROL SYSTEMS techniques to analyse, what can we do design these control
1) What does it mean to be nonlinear systems?
A nonlinear system is one where the principal of
superposition does not apply[1]. This presents a problem as III. EVOLUTIONARY TECHNIQUES
the traditional methods for analysis and design of control A. Overview
systems cannot be used. There are many examples in a wide
Genetic Algorithms (GAs) were first developed by John
range of areas where we find nonlinear systems for which we
Holland in the 1960s [10]. GAs find optimal solutions by
wish to design some sort of control system.
emulating how the process of evolution works in nature. By
using a combination of random elements, application of fitness
functions and the recombination of more successful solutions,
Manuscript submitted June 3, 201. GAs are able to ‘explore’ the possible solutions of a problem
-RH%ORJJV is a Masters of Engineering student at the School of Engineering,
Edith Cowan University, Joondalup WA 6027, Australia. (e-mail: to find optimal solutions.
MEORJJV@our.ecu.edu.au).
Using Evolutionary Techniques to Design Control System For Non-linear Systems 2

As more research occurred in this area, GAs were expanded and complex solution spaces to find optimal solutions.
upon, resulting a proliferation of terms: evolutionary Within a solution space there may be several combinations
programming [9] and genetic programming [6] are just two of that produce results that are considerably better than other
the terms used. However, all these techniques feature some neighbouring solutions (those solutions that have similar, but
common elements. not identical, parameter values). If a solution’s parameters are
varied away from these local maxima, its fitness value would
B. Definitions
start to decrease, indicating that the searcher should head back
1) Solutions towards to local maxima. However, these local maxima are
A solution to a problem is a set of values for the available not the global optimal points and can present a problem for
parameters of the system. For example, for a circuit design, any optimisation algorithm. Any search technique needs to be
these would be the resistances, capacitances and inductances able to take into account local maxima and present a way to
of the various components, plus information on how each take leaps away from them in an attempt to find a truly global
component is connected to the other components. In a maxima.
chemical system they would be the flow rates, volumes of
tanks and their interconnections.
For each problem, there are many possible combinations of
parameters that could be used to solve that problem. Some 4
combinations of parameters will produce bad results (e.g.
blow up the tanks or produce a non-functioning circuit), some 3
will produce good or even optimal results and some solutions
2
will sit somewhere in between (i.e. they may produce the
desired outcome but cost more to build or they may produce 1
useful but sub-optimal results). 2.2
0 1.1
C. Assessing Solutions and Fitness Functions
0
How do we tell if one solution is better than another? In ‐1 0 0.4 0.8 1.2 1.6
2 2.4 2.8
control system we look at criteria such as stability, steady-
state error, delay time, rise time, peak time, maximum ‐2
overshoot, settling time, etc [1]. We may also be concerned
with other factors such as cost to build, number of components
and size. To assess the suitability of one solution over Figure 1: A sample solution space.
The fitness function takes in two values and generates a score for the
another, we need to define one or more ‘fitness functions’ to fitness, measured by the vertical axis. This system has multiple peaks
measure how well a given solution meets our desired criteria. where there are locally maximal solutions, but there is only one global
A fitness function will produce a numerical value reflecting optimal solution.
how well a solution meets the desired criteria and may
combine several factors; for example a control system’s D. Components of a ETs and design decisions
fitness function could be a combination of the steady-state 1) Choice of parameters and define the problem.
error, maximum overshoot and cost to build of the system, The initial step of applying ETs to a problem is to define
with scaling factors to ensure that all three values contribute what parameters you are trying to optimise. For some control
equally to the fitness value. system designs, the developers have already decided upon the
Developing a fitness function can be quite tricky, especially configuration of the controller they are going to use (for
when dealing with opposing objectives (such as cost verses example a PID controller [9] or a sliding mode controller [11])
performance) or orthogonal objectives (performance verse and just wish to use the ETs to develop the parameters of the
aesthetics). controller. A more complex approach is to define the ET in
1) Solution Space such a way as to allow it to control the interconnection of
The range of possible solutions is known as the ‘solution elements, evolving the control equations [6] or to help with
space’ of the problem. Somewhere in this multi-dimensional design of other advanced control systems such as neural
solution space lays the optimal solution for the problem, as networks [2] or fuzzy logic controllers [12].
defined by our fitness function(s). 2) Encoding the Solutions
If the fitness function is relatively simply, we have many For an ET to work, it requires the potential solutions to be
techniques available to find where the optimal solution will expressed in terms of a number of ‘genes’. This encoding
lie. For instance, with the example shown in Figure 1, simple scheme is required to be developed to allow for the solution to
differentiation would enable us to determine the optimal be divided into components for use in the ‘breeding’ stage of
solution. However, when we encounter more complex fitness the algorithm (see below). The choice of how a solution is
functions, we need different techniques to sift through all the encoded will depend on the parameters that have been chosen
possible solutions to find the good ones. Evolutionary to be optimised.
techniques offer one such method of searching through large The simplest method for a set of numeric values is to simply
express those numbers in binary and concatenate them into a
Using Evolutionary Techniques to Design Control System For Non-linear Systems 3

single value. However, this may not be the best way to combining the features of the fitter solutions, the system will
structure the encoding in all problems. For example, when eventually produce solutions that combine the best
using ETs to develop the control for a UAV (Unmanned Arial characteristics of all the solutions present.
Vehicle), Bourmistrova and Khantsis [6] needed to design the
encoding to be able to express equations in a meaningful way c) Mutation
as they were not just selecting parameters, but how those To allow for parameter values that were not present in the
parameters are combined. original population, random changes known as mutations are
3) Evaluating Fitness applied to solutions as part of the process of creating a new
An important part of the operations of an ET is the generation of solutions. Applying mutations also helps to
evaluation of the fitness of the encoded solutions. Commonly, prevent the population becoming stuck on a local maxima.
the parameters of a solution are feed into a single fitness However, the frequency and degree of mutation needs to be
function and a single numeric value is returned. However, carefully set, as too frequent or too large mutations can
sometimes multiple fitness functions are used, such as in the destroy the work the fitness based breeding operations have on
co-operative co-evolution genetic algorithm (CCGA) [13]. improving the overall fitness of the population and the
The choice of the fitness function(s) is important, as it dictates progress towards optimal solutions.
what goals the system will be optimizing towards. 5) Problems with Evolutionary Systems as Optimisation
For a given fitness function or functions, there may be Techniques
multiple optimal solutions. For functions with multiple While ETs are powerful optimisation methods, they do have
unrelated goals, there will be some solutions that balance out a number of disadvantages. Firstly, there are the design
the various goals, while others may favour a single goal to the challenges of coming up with the correct encoding scheme for
detriment of others. Various different ET strategies try to the solution, designing the fitness function(s) and selecting
compensate for this problem and are important considerations the correct parameters for the population size, breeding
when designing a program to solve a particular problem. functions and mutation functions. Making the correct
4) Operations and Iterations decisions in these areas is important to the success of the
ETs work over a large number of iterations, known as optimisation process.
generations. For each generation, the ET will perform a Another limitation of ETs is that they can be computational
number of operations to produce a subsequent generation. expensive to run (having to evaluation fitness for a large
With each generation, the ET should be progressing towards number of solutions per generation) and require many
the optimal solution(s). generations to successfully find an optimal solution.
Furthermore, as they rely on random elements as part of
a) Initial population and population size several operations, there is no definitive way to determine how
At the start of the process, an initial selection of solutions long (if every) it will take before they find the optimal
needs to be generated. This is done by using a random solutions to a problem. These two time elements can restrict
number generator to create a number of random solutions. At the use of ETs if a rapid answer to a design decision is
this stage it is unlikely that any solution is optimal. required.
The number of solutions in the population is one of the
factors that can be modified to affect the outcome of the ET. IV. EXAMPLES OF APPLYING EVOLUTIONARY TECHNIQUES
Too few solutions or two many can both have detrimental Evolutionary techniques have been applied in the
effects on the outcome. development of control systems for many non-linear systems.
Below are four examples.
b) Breeding
To produce the next generation of solutions from the current A. Controlling an Oil Tanker
one, a ‘breeding’ method is used. Elements from two (or McGookin et al [11] used a genetic algorithm to design a
more) solutions are combined together to form a new solution. Sliding Mode controller for the steering of an oil tanker.
This can be as simple as selecting a random point in the binary Using linear models for control of oil tankers has difficulties
array representing the solutions and ‘cutting’ the solution at as their overall performance changes significantly with the
that point and combining it the complementary section from speed of the vessel and the depth of the water it is in. As such
another solution, to more complex combination techniques. non-linear controllers, such as Sliding Mode controllers, are
The most important part of the breeding process is that desirable. However, they present difficulties in selecting the
solutions are chosen to contribute to a new generation based parameters of the controller.
on their fitness function results. Solutions that have higher McGookin et al designed their GA to optimise 4 parameters
fitness values should be selected more frequently to contribute – the first and second closed loop poles, the switching gain
to the next generation of solutions. Low fitness solutions and the boundary layer thickness of the heading controller.
should contribute to few or maybe no new solutions. Their fitness function was the weighted sum of the error
Over the generations of solutions, those with elements that between the desired and obtained heading and the rudder
produce higher fitness values will become more common deflection. Rudder deflection was included as they wished to
within the population. By producing new solutions by minimise the movement of the rudder.
Using Evolutionary Techniques to Design Control System For Non-linear Systems 4

The solution that the GA produced was compared against a density, is fuel flow valve position, is steam control
manually tuned solution for the problem and was found to be valve position and is feedwater valve position. The valves
slightly more optimal, with the advantage that it was able to all have limitations on their rate of change and maximum and
find this solution without any prior information about the minimum values.
approximate region of the optimal solution and in much Dimeo and Lee used genetic algorithms to design a coupled
quicker time. PI controller and a linear-quadratic regulator (LQR) controller.
To determine the parameters for each system, they used
B. Boeing 747 Autopilot Control
different population sizes and parameters for developing a
Manocha [4] models the four forces that apply to a aircraft: generation of solutions. They ran their GAs until either 30
thrust, weight, drag and lift. These forces result in linear and generations had passed, or until they had 95% convergence on
angular motion about the x, y and z axis, and are a set of non- a single solution in the population. They found that the PI
linear equations. By working through these equations and controller they developed had good steady-state results, but
using data for a Boeing 747-400, Manocha developed the some strong oscillations during the transitive response. Their
following transfer functions for pitch control, roll control and LQR controller preformed well, but has a small steady state
yaw control: error.
1.312 0.3896
 
 0.70 2.41 0.3896 V. CONCLUSION
(5)
When working ‘in the real world’, an engineering cannot
Φ 0.1732s 0.03999 0.155
  expect to find problems to neatly fit into the introductory
s  0.7387 1.205 0.4364 0.155 topics covered in an undergraduate course. It is important for
(6) a control engineer to be able to deal with nonlinear systems, as
Ψ 0.051s  0.022 there many real world problems that may require a control
 
s  0.7387 1.09 0.4488 system that cannot be adequately expressed as linear
(7) equations. These problems arise across all areas of
Manocha then used MATLAB’s GA library to determine engineering.
the parameters for the pitch, roll and yaw PID control systems. Engineers often look to nature for inspiration in how to
These values were then entered into autopilot software in a solve design challenges. As such, it shouldn’t be surprising
flight simulator for testing. Manocha described the results that evolutionary techniques can be successful used to solve
from the GA as producing an ‘enhanced autopilot system’. many control system problems.
C. Modelling a Twin Rotor System Evolutionary techniques such as genetic algorithms offer a
control engineer a useful tool for designing controllers for
Darus, Aldebrez and Tokh [5] compared the use of a GA non-linear systems. ETs can also be used where the exact
against a recursive least squares (RLS) method for modelling a design is unknown, rather than the component values for a
twin rotor multi-input multi-output system (TRMS). The design. While harder to construct, these ETs could help were
TRMS rig used consisted of a beam on a pivot that could a starting design is only vaguely known, or when certain
freely rotate horizontally and vertical. At one end was a sections can only be treated as complex ‘black-boxes’. It is
horizontal rotor to produce lift while at the other end a smaller also important to understand the drawbacks associated with
vertical rotor was used to make the beam turn left or right. ETs, namely time taken to run them and the sensitivity to
The system is highly nonlinear. implantation choices that can have a dramatic impact on
The GA used had 100 solutions in each generation and was results.
run for 500 generations. The model obtained from the GA
was compared against the model derived using RLS and REFERENCES
against the physical model. The GA derived model was had
[1] K. Ogata, Modern control engineering, Fifth ed. Boston: Pearson,
approximately half the error as the RLS model and was within 2010.
95% confidence level of the physical model. [2] N. Metni, "Neuro-control of an inverted pendulum using Genetic
Algorithm," in Advances in Computational Tools for Engineering
D. Power Plant Applications, 2009. ACTEA '09. International Conference on,
2009, pp. 27-33.
Dimeo and Lee investigated the use of a genetic algorithm [3] K. SULTAN. (2003). Inverted Pendulum: Analysis, Design and
to control a Swedish boiler-turbine plant [8]. The system has Implmentation. Available:
the following characteristic equations: http://www.mathworks.com/matlabcentral/fileexchange/3790
[4] A. Manocha and A. Sharma, "Three Axis Aircraft Autopilot
0.0018  0.9  0.15 Control Using Genetic Algorithms : An Experimental Study," in
Advance Computing Conference, 2009. IACC 2009. IEEE
International, 2009, pp. 171-174.
0.073 0.016  0.1 [5] I. Z. M. Darus, et al., "Parametric modelling of a twin rotor system
using genetic algorithms," in Control, Communications and Signal
141   1.1 0.19 Processing, 2004. First International Symposium on, 2004, pp.
85 115-118.
where is drum pressure, is power output, is fluid [6] A. Bourmistrova and S. Khantis, "Control system design
optimisation via genetic programming," in Evolutionary
Using Evolutionary Techniques to Design Control System For Non-linear Systems 5

Computation, 2007. CEC 2007. IEEE Congress on, 2007, pp.


1993-2000.
[7] X. Peng, et al., "Research and Application on GA-Based Two-
Stage Fuzzy Temperature Control System for a Type of Industrial
Furnace," in Electrical and Control Engineering (ICECE), 2010
International Conference on, 2010, pp. 1558-1561.
[8] R. Dimeo and K. Y. Lee, "The use of a genetic algorithm in power
plant control system design," in Decision and Control, 1995.,
Proceedings of the 34th IEEE Conference on, 1995, pp. 737-742
vol.1.
[9] Y. Xiao and J. Jiang, "Evolutionary programming in non-linear
control systems," in Industrial Electronics, 1999. ISIE '99.
Proceedings of the IEEE International Symposium on, 1999, pp.
1290-1293 vol.3.
[10] Perspectives on Adaptation in Natural and Artificial Systems :
Essays in Honor of John Holland. Cary, NC, USA: Oxford
University Press, 2005.
[11] E. W. McGookin, et al., "Non-linear tanker control system
parameter optimisation using genetic algorithms," in OCEANS '97.
MTS/IEEE Conference Proceedings, 1997, pp. 17-22 vol.1.
[12] L. Joon-Yong, et al., "Multi-objective evolutionary design of fuzzy
logic controllers for car parking problem," in Advanced Motion
Control, 2006. 9th IEEE International Workshop on, 2006, pp.
607-612.
[13] N. Chaiyaratana and A. M. S. Zalzala, "Recent developments in
evolutionary and genetic algorithms: theory and applications," in
Genetic Algorithms in Engineering Systems: Innovations and
Applications, 1997. GALESIA 97. Second International
Conference On (Conf. Publ. No. 446), 1997, pp. 270-277.

-RH is currently aQHQJLQHHULQJ consultant to the 3DWHQWV2IILFH


of Western Australia, and is studying for his Masters of
Engineering at Edith Cowan University, Western Australia.

Вам также может понравиться