Professional Documents
Culture Documents
2002
EN9940
Memorial University
Page 1 of 20
2002
Summary
This paper gives an introduction to the basic concepts of classical control systems and intelligent control systems. Further it contrasts the two control system areas and gives examples of how each would be applied.
EN9940
Memorial University
Page 2 of 20
2002
Table of Contents
Summary ............................................................................................................................. 2 Table of Contents ................................................................................................................ 3 List Of Figures .................................................................................................................... 4
1. Introduction ..................................................................................................... 5
1.1 Distinguishing between Classical and Intelligent Control ........................................ 5 1.2 Shifting Intelligence .................................................................................................. 5
EN9940
Memorial University
Page 3 of 20
2002
List Of Figures
Figure 1. Shifting Intelligence............................................................................................ 6 Figure 2. An Open-Loop Control System. ......................................................................... 7 Figure 3. A Closed-Loop Control System.......................................................................... 8 Figure 4. First Order and Second Order Transfer Functions.............................................. 9 Figure 5. Model of a Cruise Control System. .................................................................... 9 Figure 6. PID Control Signal............................................................................................ 10 Figure 7. Fuzzy Logic Control of the distance between two cars. ................................... 12 Figure 8. Fuzzy subset intervals for Distance, Velocity and Braking Force.................... 12 Figure 9. Biological neurons interconnected.................................................................... 13 Figure 10. A single artificial neuron showing its parts. ................................................... 14 Figure 11. A Backpropagation-ANN. .............................................................................. 15 Figure 12. Backpropagation Algorithm ........................................................................... 15 Figure 13. Mutation (left) and Crossover Operations in Genetic Programming.............. 17
EN9940
Memorial University
Page 4 of 20
2002
1. Introduction
Control systems theory [1, 9] has always been at the heart of Robotics. Whether it is a mobile robot or a robot arm, in general, some sort of control systems theory is required. Two simple examples are easily realizable in the Intelligent Systems Laboratory (ISLAB). First, the RT200 robot arm utilizes a control system to move each arm segment to its setpoint. Secondly, the wheelchair robot utilizes a control system to move itself around. Over the years, many methods for control have been developed. Each of these has had its own advantages and disadvantages over other methods. The different methods of control can be divided into two categories: Classical Control (CC) and Intelligent Control (IC). The differences between these two categories are explained below.
EN9940
Memorial University
Page 5 of 20
2002
EN9940
Memorial University
Page 6 of 20
2002
The important thing to note in the above figure is that the output of the process (ie the plant) isnt fed back to the controller in any way. An open-loop controller (OLC) applies a correction effort when needed and assumes the desired results will be obtained. As an example, an oil pipeline may have an emergency shutdown controller that automatically shuts the system down and clears the pipes if a spark is detected.
EN9940
Memorial University
Page 7 of 20
2002
The important thing to note in the above figure is that the output (o) of the process (ie the plant) is fed back to the controller in the form of a difference in the desired signal (i o). A closed-loop controller (CLC) can keep a cars speed constant through an uphill climb and keep an ovens temperature within a range. Although every CLC has its own strategy for accomplishing its goal, the underlying idea is the same: measure a variable, if it is not acceptable, apply a correction process and repeat. It is interesting to note that CLCs may sometimes act as OLCs. For example, a sensor may not provide a feedback signal or an operator may take over the feedback operation to manipulate the controllers output manually.
EN9940
Memorial University
Page 8 of 20
2002
EN9940
Memorial University
Page 9 of 20
2002
where u is the force from the engine, m is the mass of the car, v is the velocity of the car and bv is the force due to friction.
EN9940
Memorial University
Page 10 of 20
2002
EN9940
Memorial University
Page 11 of 20
2002
Second, define the subsets intervals. To simplify things, 3 subset intervals will be chosen for each variable. These are low, medium and high for distance and velocity, and small, medium and big for braking force. These subset intervals are illustrated in Figure 8.
Figure 8. Fuzzy subset intervals for Distance, Velocity and Braking Force.
Third, choose the membership functions. In this example, the shape of the membership functions are fairly simple, just a linear transition between the various subsets. See Figure 8. To illustrate, in Figure 8 the membership function for low distance goes down linearly from 1 to 0 as distance goes from 0 to 5 meters. Fourth, set the IF-THEN rules. This is how combinations of the input determine the output. For example, IF the distance, D, between the cars is low AND the velocity of the following car is high, THEN the braking to apply is big. Similarly, the other rules are defined. This is where the non-quantitative human reasoning comes in. Fifth, perform calculations and adjust the rules. Since the rules are non-exact some adjustments may be necessary to more optimally control the vehicles distance. As an example, say the distance between the vehicles is 2.5 meters and the speed of the following car is 100 km/h. From the distance subset in Figure 8, the 2.5 meter distance translates into 0.25 medium distance plus 0.5 low distance. Likewise, the 100 km/h speed translates into 0.75 high speed. Various methods to determine an output can be EN9940 Memorial University Page 12 of 20
2002
applied. Here the minimum of the two possible subset membership of the two inputs will be used. A 0.25 medium distance AND 0.75 high speed gives 0.25 medium braking. A 0.5 low distance AND 0.75 high speed gives 0.5 big braking. The center of gravity of the area under the braking curve due to these two portions is computed and taken as the amount of braking to apply to the following vehicles. Although the center of gravity was used in this example, many other techniques exist for determining the output.
ANNs act similarly. Artificial neurons receive input from other neurons through a weighting function. See Figure 10. This is usually an amplification or suppression of the signal. All of such signals connected to the neuron are added together. If this sum is higher than some threshold the neuron will fire and send out its own signal to other EN9940 Memorial University Page 13 of 20
2002
neurons. The output of the neuron is often determined by a sigmoid function of its input rather than the threshold function. This gives the neuron a non-linear input to output relationship. It should be noted that the knowledge is stored in the input weights of the
neuron. Adjusting these weights give the neurons the ability to store different information. One neuron cant store much information, but many neurons, interconnected in several layers can store much information. See Figure 11.
EN9940
Memorial University
Page 14 of 20
2002
EN9940
Memorial University
Page 15 of 20
2002
In the above figure, ok is the output for layer k, whi is the weights from node i to h, xi is the input from node i, is the neural function (threshold or sigmoid), is a measure of how much a neuron should be adjusted by, t is what the output of the ANN should have been, and is the learning rate. This method of developing ANNs has been around for some time, and has found many uses. The most obvious examples would be to use an ANN for the navigation and obstacle avoidance of a mobile robot [3], and to use an ANN as a servo controller for controlling robotic manipulators [3].
EN9940
Memorial University
Page 16 of 20
2002
The above three steps are repeated a set number of times or until a desired fitness value is reached. Which ever program works the best is the result of GP.
2002
based on these choices. These rewards/punishments affect how the agent will make its choices when it encounters the same scenario again. Thus it learns what choices are rewarding and which are not. There are many examples of uses of reinforcement learning in an earlier paper for this course [2]. The reader is referred there for an in-depth study of reinforcement learning. Some of the examples include the inverted pendulum problem and control of a walking hexapod.
EN9940
Memorial University
Page 18 of 20
2002
4. Conclusion
Classical control systems and intelligent control systems both have their place in control systems theory. Usually when classical control can be used, it is. This is due to the simple fact that intelligent control can be over-kill in some situations. Intelligent control systems excel in areas that are highly non-linear, or when classical control systems fail, or a model of the system is difficult or impossible to obtain. Robotics lends itself to the use of intelligent control due to the fact that the system is difficult (often impossible) to model. In fact, the non-deterministic nature of robotics almost requires some sort of intelligent control in order to be effective.
EN9940
Memorial University
Page 19 of 20
2002
5. References
[1] Phillips, C. L., Harbor, R. D., Feedback Control Systems, 2nd Ed., Prentice-Hall, 1991 [2] Seshadri, J., Reinforcement Learning, ISL-0008 [3] Morris, A. S., Zalzala, A. M., Neural Networks for Robotic Control: Theory and Applications, Ellis Horwood, 1996 [4] Smith, B. Support Vector Machines for Object Recognition, NECEC 2001 [5] Scholkopf, B., Burges, C., Smola, A., Advances in Kernel Methods: Support Vector Learning, The MIT Press, 1999 [6] Suykens, J. A. K., Vandewalle, J., De Moor, B., Optimal Control by Least Squares Support Vector Machines, Neural Networks 14, pp 23-35, 2001 [7] Koza, J., Genetic Programming, http://www.geneticprogramming.com/johnkoza.html [8] Bauer, P., Nouak, S., Winkler, R., A brief course in Fuzzy Logic and Fuzzy Control, http://www.flll.uni-linz.ac.at/pdw/fuzzy/fuzzy.html [9] Nise, N., Control Systems Engineering, 2nd Ed., BCPC, 1995
EN9940
Memorial University
Page 20 of 20