Focus more on theoretical function

Today, I tried several algorithms and adjust some parameters of ANN to improve the model. But R2 still did not improve much (about 0.5<R<0.6).

Summary:

The following may be the best model I get today.
The first are comparisons of predictions and targets.

The second is all 3042 first parameters.

The third is log10-scale plot of y-axis of 3042 first parameters.

 The fourth is the list of algorithms.

Tomorrow, I will try to look into these algorithms to find the best one for my research.

improve ANN a little

Today,  I checked all the data of 6 parameters. I found that there are some very strange numbers. They are either very large (more than 1000 times higher than the mean value) or very small (negatively large). So I deleted them ( 13 in total). And build a slightly better ann model.

Summary:

The following are 6 parameters' summaries after deleting strange numbers.

> summary(betachange3$P1)

     Min.   1st Qu.    Median      Mean   3rd Qu.      Max. 

0.0000000 0.0006518 0.0013720 0.0015680 0.0021490 0.0204000 

> summary(betachange3$P2)

   Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 

-98.010   5.283  12.250  13.450  20.900  46.640 

> summary(betachange3$P3)

   Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 

 -9.448   2.174   3.447   3.539   4.808  39.760 

> summary(betachange3$P4)

      Min.    1st Qu.     Median       Mean    3rd Qu.       Max. 

-0.0001874  0.0007157  0.0012470  0.0016700  0.0021710  0.0113300 

> summary(betachange3$P5)

   Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 

  1.479  35.100  42.320  40.240  47.710  76.590 

> summary(betachange3$P6)

   Min. 1st Qu.  Median    Mean 3rd Qu.    Max. 

 -2.535   1.875   2.513   3.056   3.706  19.050 

The first are the comparisons of outputs and targets.

The second is all the 3042 predicted outputs (red) and original targets (blue).

The third is the first 500 pairs.

Although it is not very accurate, it improves a lot compared with the original models.

Next week, I will continue to improve them by other ways.

read one paper and think of two ways to improve models

Today, I read one paper and thought of two way to improve models.

Summary:

First, for ANN, if I can set two or more layers of the fitting model, it may be more accurate to get the results. So I searched online and find how to set two or more layers. The codes are still running, I will update my blog tomorrow to show the results.

Second, as we discussed, I tried to divide the reservoir into different formations to at least prove that the model is suitable.

Although the prediction result is not very good, at least they are on the same scale. I will try to improve it tomorrow too.

Tomorrow, I will look more into ANN model and SVM model.

2 new SVM models

Today, I learned how to build SVM models and built 2 SVM models.

Summary;

first model:

second model:

The first figures for both models are outputs of the regression with the minimum estimated cross-validation loss.
The second figures for both models are comparison of observation values and prediction values, which do not show good results.

Tomorrow, I will try to improve the models to get better results.

try ANN and SVM models for logging data and T2 parameters

Today, I tried ANN and SVM again.

Summary:

ANN:

SVM:

svmstd =

  RegressionSVM
             ResponseName: 'Y'
    CategoricalPredictors: []
        ResponseTransform: 'none'
                    Alpha: [1104×1 double]
                     Bias: 0.0015
         KernelParameters: [1×1 struct]
                       Mu: [1×16 double]
                    Sigma: [1×16 double]
          NumObservations: 3055
           BoxConstraints: [3055×1 double]
          ConvergenceInfo: [1×1 struct]
          IsSupportVector: [3055×1 logical]
                   Solver: 'SMO'

The result is like that. The model converged after 552 iterations.
The MSE of the model is 1.1725e-6.

The model is not good enough yet. There may be two reasons.
First, for ANN, 3055 observations with 16 predictors may not be big enough to build a good model. The data matrix is not big enough. So ANN is not suitable for building the model.
Second, We should delete some noises of logging data manually.

Tomorrow, I will continue to apply it to SVM and verify two assumptions mentioned above.

try several models

Today, I tried several models. They include Cubist, Random Forest and SVM. But after trials and errors, I could not find that they have good modeling results.

Summary:

SVM:

 Random Forest:

Cubist:

All of their R2 are lower than 0.5, which show bad modeling results.

Yesterday's ANN flow chart:

I find that there are also some packages of ANN and other models in Matlab. 

Next week, I plan to look into them to find if they can be helpful for my research.

Draw a flow chart for ANN and draw the ANN prediction results

Today, I draw a flow chart for ANN and draw the ANN prediction results, which is not good.

Summary:

I sent you a ppt file to show the flow chart of how to build an ANN model.

The following is the bad behavior of this ANN model.

Y is our data. X is predicted data. They should be near the line (y=x).

Tomorrow, I will try to improve the model or find another better model to predict our data.