It appears that neural networks can be created more efficiently than evolution has done, even though evolution favours energy efficiency. We should be able to obtain activity in an ANN similar to that of a natural neural network more efficent than NN, if it is carefully planned . Intelligent design IS possible for scientists.

How large is your NN?

Small ones are good at solving toy-size problems, but the training time and the efficacy quickly get worse as the problem scales up.

"Neural networks are the second best way of doing just about anything."

John Denker

FNN,TDNN,MNN, .... Lots of work from the last 10-20 years in NN controllers.

Most important is the learning of dynamics.

contact me if you are interested: [email protected]

Lots of work, but I do not find it very convincing.

Without clever reduction by NNNs (natural neural networks, in skulls) to a small problem,

[alphabetSoup]NNs continue to scale badly.

By historical input signals you help MLP network to learn dynamic of control process but the structure of network is the same and there is just extra inputs. And RNN networks because of inner feedbacks ,some times can learn to control without historical inputs.

Tim is right. Very hard to work well and cross validation problem.

I would rather use Fuzzy Logic, very much faster in practice, or ANFIS (Yang 1993) when you have a lot of inputs/outputs data.

Dear Dinie Muhammad,

Artificial neuron network is a toy ad infinitum. The more elements in a neuron network then the greater a variety of the connections and there are more persons, which will add one more feedforward or feedback for the decision of the problems of the controlling or classification. Through 50 years all this will leave in a history, as the perceptrons. Certainly, they excited an idea.

The nervous system works completely by on other principles.

My approach consist in studying nervous system through creation of its model. See, please: Romanov S.P. Modelling of ionic channel properties and its role in shaping neuronal spike activity // Neurophysiology. 1989. V. 21. N 3, p.277-285.Romanov S.P. Servo-loops as homeostatic mechanism of motor control on segmental level / Posture and Gait: Control Mechanisms, 1992. XI-th International Symposium of the Society for Postural and Gait Research, Portland, May 24-27 1992. Vol.1. Eds. Marjorie Woollacott and Fay Horak, University of Oregon Books, 1992, p. 71-74.Romanov S.P. Information processes in neuron network of muscle contraction control // Vestibular and neural front / Proceedings of the 12th International Symposium on Posture and Gait, Matsumoto, Japan, 3-7 October 1994. Elsevier Science B.V. Amsterdam, The Netherlands, 1994. P. 457-460. Romanov S.P. Proprioceptive feedback as a homeostatic mechanism of multiparametric motor control // Motor Control VII. Proceedings of the VIIth International Symposium on Motor Control held in Borovets, Bulgaria, June 21-25, 1993 / Douglas G. Stuart (ed.). Motor Control Press: Tucson, AZ. USA. 1996, p. 141-145.

To my regret, all from basic my works in Russian.

And now I move at my work to the higher levels of motor control system.

My best regards.

Thank you,

Sergey Romanov

Dinie,

the models that you present in the figure not considers the error, comparing the ref and the output system. try to train the ANN off line with samples from control tests by other kind of controllers like PID ou Fuzzy to a close range of operation of the system, and after this, use the mathematical model of ANN replacing the PID by its. Dont forget to use a error as input of your ANN controller.

Adelson - > [email protected]

Please read:

http://www.researchgate.net/publication/216774240_Artificial_Interactive_Neural_Network_(AINN)?ev=prf_pub_p2

Article Artificial Interactive Neural Network (AINN)

you read this book

Neural Networks Algorithms, Applications, and Programming Techniques

James A. Freeman

David M. Skapura

Dear all,

Thank you very much for you comments, opinions and reference materials. I really appreciated it.

My project is actually to control the distillation column using Nonlinear Internal Model Control (NIMC). Thus, the model and the controller is developed using NN. Since the process/system (distillation column) is developed from Aspen Dynamic (a process simulator software), I have many data. Yet, the setback is the process have a slow dynamic which I end up have to run the simulation (during data generation) for a long time just to get some pattern of the process for training. So, the my training sample is quite big. For the controller, I just use 3 inputs and 1 output which is r(k+1), y(k) and y(k) to predict u(k).

Well, after having some not-so-good experiences with NN, I do agree with Tim Poston and Sebastien Dourlens. In my opinion, NN is have many unresolved issues making it not very promising especially as a controller. Develop the NN using heuristic approach takes a lot of time and the result is unpredictable. Yet, I do not have much luxury to learn a new method (e.g ANFIS and Rita Fizzi SOM ) since my project deadline is near. So, I have to make the best of what I have now.

@ Mehdi The problem comes when I want to apply it online (in Matlab Simulink). Even though during the training of FNN, I had use the u(k-1) as input, but when I apply it on Simulink that u(k-1) input is the historical input from NN controller (no longer from u(k-1) training data) . So, using the same FNN will produce a poor result because when training, I used historical input from the controller (training data) but when online application the same input had to be represent by NN controller output. (I had try my best to explain the problem. Sorry if its still unclear).

There are many paper regarding NN controller using FNN but rather in RNN scheme (from my perspective; train like FNN but implement like RNN). But, from my 'novice experience', I notice that all did not bring out this issues. Or it just only me have this problem? An example is the attached figure. The researcher stated that they are using FNN but as we can see in the figure the scheme is suit better to RNN (for controller). Here is paper link;

http://www.informaworld.com/10.1080/00986449908912768

I had this dilemma for quite some time. I really apprenticed if someone can help me.

Thank you in advance.

@Adelson In my opinion, the IMC scheme is a bit different from PID. The input of the controller is the modified setpoint not the error between the setpoint and the system response. Or is better for me to include the both (error and modified setpoint) ? Regarding your other comment, if I train the NN like that, wouldn't it will be just mimicking the PID or Fuzzy response of the system?

the answer to the second question is: yes, its is just mimicking the other controllers, but if you train with a significant sample, the ANN controller will generalize to the others situation, but this is just one of al ways possible.

the answer to the first comment i think that always in feedback control systems the error is the input of the controller even in predictive control like yours.

I disagree with almost all your statements. Neural networks are very efficient if you do them right. No magic, almost no longer any "open issues", just maths, statistics, and a lot of automatic control if that's what you are doing; no biology at all, of course, despite the name. Read good textbooks about neural network control, i.e. textbooks written by experts in automatic control (in the above comments, nobody seems to be concerned with noise and disturbances, which are key factors). Another key issue is the quality of your model; as very rightly pointed out by Tim Poston, use your brain, i.e. your knowledge about the distillation process, which will allow you to build a parsimonious neural-network based semi-physical model without serious scaling problems (see http://www.neurones.espci.fr/Articles_PS/EARLY_FAULT.pdf for an application of semi-physical neural modeling to a distillation column, see http://www.neurones.espci.fr/Francais.Docs/dossier_recherche/Articles_PS/GRAYBOX.pdf

for a general introduction to semi-physical neural modeling, and see http://www.neurones.espci.fr/Articles_PS/rivals2.pdf for an application of semi-physical neural modeling to the control of a real 4WD vehicle). I just hope you are not a novice in neural networks, in automatic control and in chemical engineering.

In addition, neural-network based industrial control modules are commercially available. I think I remember Emerson offering a module called DeltaV. Siemens has also been very active in neural control for many years and has implemented many industrial applications.

" The models that you present in the figure [do] not consider[] the error, comparing the ref and the output system. try to train the ANN off line with samples from control tests by other kind of controllers ...."

I worked from the assumption that what is fed back into the input layers DOES indeed include the error. If not, then training offline is likely needed.

But let's assume you do include the error in the input fed back into the input layer of a 'vanilla' perceptron with one hidden layer. I recall an architecture called the 'cascade correlator' that started with a hidden layer of one [or even zero?] hidden layers that added one hidden unit at a time as long as the error variance decreases BUT STOPS AS SOON AS THE IMPROVEMENT STOPS OR BECOMES SMALL RELATIVE TO THE PREVIOUS REDUCTIONS. That way you essentially get a principle components analysis of the error stopping the increase in dimensions at the point of diminishing return.

Did I ever try it? well---no [blush]

Hi,

I don't know what methods people from the control theory field usually use, but control problems can be addressed using methods from the field of reinforcement learning.

I think that artifical neural network may control the behavior of activity of neurons from hidden layers. The system of input layer and output will influence feedforward neural network-based modeling on classical neural networks, for example, PDP model, M-P model, B-P model, Connectionist model etc. Thereby FNN ought to consider these problems to occur.

It appears that neural networks can be created more efficiently than evolution has done, even though evolution favours energy efficiency. We should be able to obtain activity in an ANN similar to that of a natural neural network more efficent than NN, if it is carefully planned . Intelligent design IS possible for scientists.

Lazar Puhalo: very good

i have used only FNN with prediction, and as you said it is very fast function to use. sorry, I did not work with system control.

hello, i have something to say about the generalization capacity of the ann. The rnn has more capacity to generalize over examples that have not presented during the training phase but their outputs have less precision that the fnn.

You can prefer RNN or use one of Model Reference controller, NN Predictive controller, NARMA-L2 Controller, If you have worked on Simulink.

Thank you all for your comments.

@Gérard Dreyfus What you said is very much true. Unfortunately, I still regard myself as a beginner in NN and control system. Thanks for the reference materials. I hope you can suggest me some good text books regarding NN control from experts in automatic control like you had mention.

@Yihong Wang What do mean by 'FNN ought to consider these problems to occur'? Do you mean like my problem?

@Wilmer Gaona Thank you for mentioning it. I will keep that in mind.

@Kamlesh Kumar Well it's true that those system is provided in Matlab Simulink. But, they did not give the source code or 'm-file' of the system. So, we do not have the capability to modified the system freely. In my opinion, it better to develop the whole system ourselves so that we really know what we are doing.

FNN like as a way of control system. However, which will challenge to the control system appearing in the classical neural network. In particular Discrete Hopfield Neural Network (DHNN) and Continuous Hopfield Neural Network (CHNN) provides the number of input equal to output from single layer network in neurons, thereby FNN provides that 'historical controller output as the input for the Neural Network controller' in control system for exploring the process of cortrolling in neural network is possible.

I think that FNN will achieve and win this working when FNN may consider the computation of input-output involved in the artificial neural network.

MNN (memory neural networks from Sastry) are far more powerful. No external loops and dynamics is very well learnt by a spatio-temporal backpropagation algorithm. One more advantage is that we can avoid any instability during learning phase. I used it in 1998 to drive a car. See my publication: Sebastien Dourlens, Nathalie Chatry, Véronique Perdereau, Michel Drouin: Lateral Position Control of a Vehicle on Highway by Memory Neuron Networks. 768-771. Neural Computation '98 . (my program runs under matlab simulink).

https://www.mathworks.com/help/deeplearning/neural-network-control-systems.html;jsessionid=cb7101cbb3f39c309452837917be

https://www.mathworks.com/help/deeplearning/ug/introduction-to-neural-network-control-systems.html

https://www.mathworks.com/help/deeplearning/ug/design-neural-network-predictive-controller-in-simulink.html