html5-img
1 / 16

Question:

Question: When we were training the anfis system, does over-fitting occur when we have a large epoch number instead of when we have large membership function number? Over-fitting =

adin
Download Presentation

Question:

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Question: When we were training the anfis system, does over-fitting occur when we have a large epoch number instead of when we have large membership function number? Over-fitting= the result when anfis automatically reduces training error in each epoch, making the system super-fits the training data, but may not fit other data in general

  2. Our best fis system is the one called “ampFisUnfiltered3.fis” which uses the following criteria: • Unfiltered data • Number of membership function=3 • Type of membership function=‘gbellmf’ • epoch_n=20

  3. Using the following code, we can estimate temperatures if we know avgAmp(average amplitude, our measured feature) • fis=readfis('ampFisUnfiltered3.fis‘) • temp_eval=evalfis(avgAmp,fis)

  4. 20100110-20100118 data plot, average amp against temperature

  5. 20100103-20100115 data plot, average amp against temperature

  6. E.g avgAmp=0.0200018485069275 temp_eval=evalfis(avgAmp,fis) >>> temp_eval=81.4416

  7. However, the true temperature is 70.7378, • We have an error of 81.4416-70.7378=10.7038 Which is considerably big, about 15% of true value. Could it be that our fis, “ampFisUnfiltered3.fis” is not a good model to predict this relationship between avgAmp and temperature? presentableData.xlsx

  8. right Skew - the mass of the distribution is concentrated on the left of the figure. It has relatively few high values Also referred to as positive skew

  9. Another thing to be noticed, exactly 105 out of the 130 big-error data points are measured under stressed condition (elevated temperature) • this might suggest: Average amplitude of chicken vocalization as a feature of stress level works well under low-stress environment, but produces wider ranges of fluctuation under high stressed ones.

  10. Changing the type of membership function • This time we change Type of membership function=‘gaussmf’ While holding other variables constant Let’s see the behavior of this new fis system: ‘ampFisUnfiltered3Gaussmf.fis’ mfComparedataUnStressed.xlsx mfComparedataStressed.xlsx

  11. The average error and standard deviation of error vary within the two fis systems • However, after calculating the Mean Square Error for both sytems, both stressed and unstressed

  12. The smaller the MSE, the better the fis system • You can see that although very similar, it’s hard to tell which one is a better parameter, “gbellmf” or “gaussmf” since their corresponding MSE value varies for stressed and unstressed • Also the errors in unstressed groups are a lot smaller than those generated by stressed groups, which again proves our point saying “Average amplitude of chicken vocalization as a feature of stress level works well under low-stress environment, but produces wider ranges of fluctuation under high stressed ones.”

  13. The end

More Related