About the NeuroFuzzy Module of the FuzzyTECH5.5 Software
Contents
1. a variable definition in which each term is determined by exactly one parameter the base value of the term s maximum As a result each term must intersect the base line at the maximum point of its neighbours On view Figure 131 in 1 3 Using the NeuroFuzzy Module Among the samples program of our FuzzyTECH edition there is a simulation program for separating the color bottles in a glass recycling plant This program can simulate the work of the human operator selecting the bottles into different containers recording the color intensity of the bottles and the type numbers of the containers where he put the bottles Using his know how we can get a training data to develop a project which can automate this process by using the NeuroFuzzy Modul In addition the simulation includes the option for testing the generated fuzzy logic system We used this simulation to learn the steps of NeuroFuzzy System Design using this modul in the fuzzyTECHS 5 online edition 3 1 Neuro Fuzzy System Design 3 1 1 Creating the Training Data Before NeuroFuzzy training we need to have the sample data in type of csv file In this case we obtained this type of training data file by initiating the simulation program in Record mode Figure shows a part of the file Nfsensor csv El Nfsensor WordPad Fajl Cel 4644 gt amp B Szerkeszt s N zet Besz r s Form tum S g Record BlueRed GreenBlue RedGreen Type Record2. obtained an initial fuzzy logic system Figure 3 fuzzyTECH 5 54j Online Edition lt untitled gt DER File Edit View Debug Tools Window Help D xc RE Bw 3k SHREW DOR SS x d s TES e H untitled t Project Editor ariable Groups E Inputs E Outputs E Intermediates XX BlueRed 5 Rule Blocks T Text ag Online Connections RB1 X GreenBlue BlueRed Type 3 GreenBlue Type Red reen Min Max X X Red reen Figure 3 The structure of generated fuzzy logic system based on Nfsensor csv Let see the rules by double clicking on the graphic element RB1 in the Project Editor window Figure 4 The rules all have the weight of zero that means there is not yet knowledge in the system The knowledge the rules for the glass sorting process gets into the system by the NeuroFuzzy training step In the column of DoS the dark backgrounds of the values display they will be learning We had to choose the CoM compromising defuzzification method because the NeuroFuzzy training uses error gradients to determine a direction of ptimization The response of the system to small parameter changes is evaluated to determine the gradients The non compromising defuzzification methods do not react to small changes in 1 p 96 HB Spreadsheet Rule Editor RB1 B umm IF BlueRed GreenBlue RedGreen low low low low low low low low low low low medium low low medium l
3. the following FDW windows you can specify everything you know about the system to design C Append to Existing System Create New System The FDW can extract information from sample data files The sample data file must be in the fuzzyTECH data format v Use a Data File Previous New End Help Cancel Figure 2 Fast designing a fuzzy logic system based on the sample data file After opening the appropriate sample data file we can see and modify the screens of Define variable Define defuzzification Define rule blocks The FDW extracts from the file a b c the number of variables the names of the variables the data range of each variable The FDW selects defaults for the fuzzy prototype system with assumptions below 1 2 3 4 5 the variables being to the last variable are inputs the last variable will be the output all inputs can be represented by three terms the output can be represented by five terms in our case we modified it to four terms default names are used for the terms terms of output variable were renamed in hungarian piros z ld feher k k 6 CoM defuzzification is used for the output 7 all variables are inferred via a single rule block and 8 acomplete rule base is created by using a DoS 0 value for all rules When we met the question Do you want to generate the specified project we answered Yes for it and we
4. 1 1739 94 1675 15 2395 96 3 Record 2 373 30 3087 05 2429 47 4 Record 3 1755 TT 1652 00 1514 98 3 Record 4 864 45 1647 31 2665 90 T Record 5 222 05 3059 54 2002 33 4 Record 6 877 88 2031 66 3071 1B 1 Record 7 1803 58 1593 12 2163 05 3 Record 68 2352 12 2557 04 1411 53 2 Record 9 401 30 3259 94 2150 96 4 Record 10 363 34 3477 95 2462 86 4 Record 11 1571 17 1731 04 1735 33 3 Record 12 104 80 3389 83 2421 83 4 Record 13 499 85 3305 75 2196 22 4 Record 14 2297 28 3340 14 535 62 2 Record 15 2092 62 SIT Els 584 32 2 Figure 1 Viewing the recorded samples in the WordPad editor The color intensities of the bottle BlueRed GreenBlue RedGreen and the type of the container in which the human operator put it Typel red Type2 green Type3 white Type4 blue to be melted for making new glass are set down in every record of the file If we had obtained the sample data we had to develop an empty initial fuzzy logic system for learning 3 1 Creating an Initial Fuzzy Logic System Based on the Sample Data Setting up a fuzzy logic system may be by Fuzzy Design Wizard FDW quickly We had to sign the Use a data file checkbox in the first dialog window Fig 2 Fuzzy Design Wizard ES Fuzzy Design Wizard Welcome to the Fuzzy Design Wizard FDW The FDW will help you to create a complete fuzzy logic system prototype in minutes By answering a number of questions in
5. About the NeuroFuzzy Module of the Fuzzy TECHS 5 Software gnes B Simon D niel Bir College of Nyiregyhaza S st i ut 31 simona nyf hu bibby freemail hu Abstract Our online edition of the software FuzzyTECHS5 5 has a NeuroFuzzy Module We studied it through the simulation sample program of the recycling glass classification We collected the experiences to develop a fuzzy logic system based on training data We need to have the sample data in type of csv file we have to develop an empty fuzzy logic system we must choose the CoM compromising defuzzification method from the offers The NeuroFuzzy modul gives us opportunity to open both rules and membership functions for learning After the last manual optimization we have a perfect fuzzy logic system 1 Introduction To expand the capabilities of fuzzy TECH the INFORM GmbH added several modules to the oline edition The NeuroFuzzy Module uses neural network technology to generate fuzzy logic rules and membership functions automatically Our online edition of this software FuzzyTECHS 54j has this module We studied it through the simulation sample program of the recycling glass classification In the next sections we describe our experiences 2 Integration the Benefits of the Fuzzy Logic and the Neural Networks Using the project editor of the program fuzzy TECH we can design a fuzzy logic system which has all the knowledge about the system including the human experienc
6. e and the fuzziness we are not able to describe by exact differential equations We can specify a fuzzy logic controlling system by a lot of rules having the form IF lt situation gt THEN lt action gt Both the lt situation gt and lt action gt are built up from linguistic variables combined with the appropriate operators To develop a fuzzy logic control system we need to take three computational steps fuzzification fuzzy inference and defuzzification We are lucky if the all knowledge about a system we want to automate we can build into the system project directly There are applications where the know how is not explicite so we need to learn it from the data set coming from the behavior of the system The neural networks are able to learn from data to set up a system In neural nets the parameters affected by the training procedure are the connection weights between the neurons Looking at a fuzzy system using arbitrary sets of fuzzy rules the firing strength of each rule output connects this rule with the precondition of the following rule block Using the technology of Fuzzy Associative Maps FAM this connection can be weighted With FAM each rule is assigned a Degree of Support DoS representing the individual importance of the rule Rules themselves can be fuzzy meaning with a support between 0 and 1 This rule plausibility can be used as the parameter to be calibrated in the neural training A neural training algor
7. ithm can calculate an error gradient by slightly changing the rule weights Integrating the explicit knowledge representation of the fuzzy logic with the learning power of the neural networks we can get the NeuroFuzzy technology The NeuroFuzzy module provides methods for supervised learning The heuristic methods used combine the two learning steps of error backpropagation with the idea of competitive learning After a system output is computed by forward propagation an output error is identified by comparing the system output with the given sample output data The standard methods perform an optimization step per sample Batch methods calculate errors and gradients for all samples and update the system s parameters after each complete iteration Standard methods provide a better computation performance and batch methods a better convergence behavior FuzzyTECH also allows us to include user defined method by writing the training method as a dynamic link library The NeuroFuzzy modul gives us opportunity to open both rules and membership functions for learning Before training is started the membership functions of all open terms of a linguistic variable are converted standard types Inputs are converted to Z Lambda and S forms For the output variables all terms are converted to Lambda type membership functions In addition the standard variables are restricted to shapes which allow an overlap of no more than two terms This leads to
8. open the learning A complete fuzzy logic system must be designed Components of the system for training must be determined A training configuration must be construct Asample file must be loaded The excellent samples and the manual help us to go forward in practice and we can test the results by the simulation program delivered also References 1 User s Manual for all fuzzyTECH 5 5 Editions 2 Andr s Botos Agnes B Simon First Steps in using of FuzzyTech5 5 Online Edition at College of Ny regyh za Proceedings of 3 International Symposium of Hungarian Reseachers on Computational Intelligence Budapest 2002 3 Simonn dr Balogh Agnes L gy sz m t stechnikai m dszerek alkalmaz s nak lehet s ge a Fuzzy TECHS 5 szoftverrel Agrarinformatikai Ny ri Egyetem s F rum G d ll 2004 4 Simonn dr Balogh Agnes Bir D niel A neur lis h l k s a fuzzy logika kombin l sa F iskol k matematika fizika s sz m t stechnika tan rainak XXVIII konferenci ja Nyiregyhaza 2004
9. ow low medium low low medium 1 2 B 4 5 5 7 8 3 low low high low low high low low high low low high low medium low low medium low low medium low low medium low low medium medium Figure 4 Rule block before learning 3 4 3 Entering A priori Knowledge The project editor allow us to make changes in the default parameters of the project enters So we modified the names of the terms Type linguistic variables from English into Hungarian 3 1 4 Opening the Components for Learning For the linguistic variables we can open specific terms for training and for the rule blocks we can open the individual rules for training by buttons marked with L letter in the small icons 3 1 5 Starting the Training To initiate NeuroFuzzy training we selected Tools Neuro Learning from the main menu of fuzzy TECH Selecting the file Nfsensor csv in the Read Example File dialog box the Learn Control dialog box opened The tool bar of this dialog box lets us control the NeuroFuzzy processing In the Figure 5 we can see the report of a learning phase Learn Control nfsensor csy b Jak 5 4 1384 60 80 Iter Deviation Values Status RandomMethod Max deviation 0 66 stop at 8 90 Exp 25 Iteration 3 Avg deviation 0 03 stopat 0 10 Time 40 ms Figure 5 Learn Control window The error plot shows the deviation of the worst sample Max and the average deviation of all samples Avg for eve
10. ry iteration The statistic plot classifies the deviations of all samples In each class the number of samples matching the deviation class is counted In our case Max deviation and Avg deviation started at 30 and lowered to 1 The percentage is calculated based on the base variable range of the output variable The training was terminated because the average deviation less then 0 05 The range of the Type is 5 The status section displays the number of the currently treated sample the number of completed iterations and the computation time for the last computation The field indicates the time for a complete iteration during continuous training and the time for the last training step while in step mode The deviation section shows the current values for the deviation of the worst sample and the average deviation The section displays different values when the NeuroFuzzy module is used in continuous training or when single training steps are taken 3 1 6 Final Optimization The result of NeuroFuzzy training 1s a fuzzy logic system that we can directly optimize by hand How to manually optimize the system depends on the application Often the objective of the training is to find functionally redundant rules or unnecessary rules Conclusions The Fuzzy TECHS5 5 program online edition is very useful program to develop an adaptiv fuzzy logic system by using the NeuroFuzzy Modul The most important steps to set up the training and
Download Pdf Manuals
Related Search