Instructions on the neural network modelling on solidification cracking. K. Ichikawa 26 September 1996 1. General All of the data and program I created or I got from david regarding the classification neural network modelling is in the my folder /users/ichikawa/ic. 2. Database "Raw" (not normalised) experimental dataset can be found in the file named /users/ichikawa/ic/hc5c without regend. Its regend can be seen in the file named /users/ichikawa/ic/tabletitle. Normalised data is in the file named /users/ichikawa/ic/inran.dat. It does not contain output (=target). We do not need to normalise targets in classification modelling (Target is only 0 or 1). 3. Final Model -Weights and Biases I created (by using all of the dataset (training+test dataset)) You can find the following files which contains the best values of weights and Biases. The 1st best model: wct02/_wk2 The 2nd best model: wct02/_wk5 The 3rd best model: wct02/_wk3 The 4th best model: wct11/_wk4 Thus my committee model requires four individual neural net model. Details of the each individual model are described in our paper. 4. Usage of the models I am explaining the procedure of a prediction using the example of the case of the prediction of complex effect of Carbon and Sulphur. 4-1. Make a input data First make a input data on which the solidification cracking should be predicted. You have to normalise the data as usual. You can see the example of the normalised data in the file; "/users/ichikawa/ic/testc3.norm". 4-2. Make a "spec" file The example can be seen in "/users/ichikawa/ic/specctestCS". 4-3. Make a "runall" file The example can be seen in "/users/ichikawa/ic/runcoCS". If you run this "runall", you need to have folders which can contain "out" and "mo" files. In my "/users/ichikawa/ic/runcoCS", I defined them as "outtestCS" and "motestCS". "outtestCS" contains "out" (Best Fit Prediction) files by the 4 individuals neural model. Similarly, "motestCS" contains "mo" (marginalised Prediction) files. If you avarage the predictions by each individual model. You can get the final committee prediction. i.e., Best Fit Committee Prediction = (out1+out2+out3+out4)/4 Marginalised Committee Prediction = (mo1+mo2+mo3+mo4)/4 However, I have already made some program to do "committee" automatically and it will be explained below. 4-4. Program "commitee.for" In the directory "/users/ichikawa/ic/s", you can find the program "committee.for". This is a program which can do a "committee" among predictions by several individual classification neural net. Details of usage of the program were written in the program as a comment line. You can see the typical example in a file "/users/ichikawa/ic/comCS.res" which contains the following output. Ave._out Min._out Max._out Ave.mo Min.mo Max.mo 1 0.00000 0.00000 0.00000 0.01350 0.00613 0.01664 2 0.00000 0.00000 0.00000 0.01308 0.00609 0.01614 3 0.00000 0.00000 0.00000 0.01284 0.00607 0.01598 4 0.00000 0.00000 0.00000 0.01267 0.00612 0.01542 5 0.00000 0.00000 0.00000 0.01247 0.00626 0.01551 6 0.00000 0.00000 0.00000 0.01226 0.00659 0.01611 7 0.00000 0.00000 0.00000 0.01232 0.00703 0.01644 8 0.00000 0.00000 0.00000 0.01268 0.00780 0.01662 9 0.00000 0.00000 0.00000 0.01282 0.00857 0.01615 10 0.00000 0.00000 0.00000 0.01233 0.00964 0.01465 11 0.00000 0.00000 0.00000 0.01156 0.00977 0.01348 ..........(Continue)......... Each colums has following meaning. Ave._out: Average of the each best fit prediction by individual model. i.e., Ave._out=(out1+out2+out3+out4)/4 Thus, this is the committee best fit prediction. Min._out: Minimum value of best fit predictionin predictions by each individual model. i.e., Min(out1,out2,out3,out4) Max._out: Maxmum value of best fit predictionin predictions by each individual model. i.e., Max(out1,out2,out3,out4) Ave._mo: Average of the each moderated prediction by individual model. i.e., Ave._out=(mo1+mo2+mo3+mo4)/4 Thus, this is the committee moderated prediction. Min._mo: Minimum value of moderated predictionin predictions by each individual model. i.e., Min(mo1,mo2,mo3,mo4) Max._mo: Maxmum value of moderated predictionin predictions by each individual model. i.e., Max(mo1,mo2,mo3,mo4) 5. Other programs I have created Since committee modelling require so laborious work because it needs handling of several models at the same time, I made some useful programs. Thus most of the work to creat committee classification model is automatically done. There are the following programs I created. Usages are written in the program. 5-1.extract.for 5-1-1.Description This program is to extract data from the numerous files generated on executing Bigback a number of times. Originally this program is created by Mike Lord, but I modefied in order to use my classification problems. 5-1-2.Location /users/ichikawa/ic/s/extract.for 5-2.oder.for 5-2-1.Description This is a program order the models created by large amount of neural net in terms of the magnitude of marginalised error. Before running this program, you have to use the above "extract.for" to get a well-organised data. 5-2-2.Location /users/ichikawa/ic/ordering/order.for 5-3.mkra.for 5-3-1.Description This program makes a "runall" file which can handle many models. 5-3-2.Location /users/ichikawa/ic/ordering/mkra.for 5-4.ordert.for 5-4-1.Description This program reorders the data by retraining using all the data set available and using the data extracted by the program extract.for. This is almost same as above "order.for" but slightly modefied. 5-4-2.Location /users/ichikawa/ic/ordering/ordert.for