Choosing the Best Inversion Result: Statistical Considerations

  • Context: Graduate 
  • Thread starter Thread starter marili
  • Start date Start date
  • Tags Tags
    Inversion
Click For Summary
SUMMARY

The discussion centers on selecting the optimal inversion result between two methods: using N data with k model parameters versus using 3*N data with k+3 model parameters. Key considerations include statistical techniques to determine if additional parameters merely fit noise rather than improve the model. The importance of having sufficient data to assess improvements in fit is emphasized, particularly in scenarios involving noisy data or black box techniques. Common sense and intuition about the model also play critical roles in decision-making.

PREREQUISITES
  • Understanding of statistical significance in model fitting
  • Familiarity with model parameters and their impact on data analysis
  • Knowledge of data sufficiency and its role in statistical modeling
  • Experience with black box modeling techniques
NEXT STEPS
  • Research statistical techniques for model comparison, such as AIC or BIC
  • Learn about overfitting and how to identify it in model parameters
  • Explore methods for assessing model performance with noisy data
  • Study the implications of data quantity on model accuracy and reliability
USEFUL FOR

Data scientists, statisticians, and machine learning practitioners seeking to optimize model performance and understand the implications of adding parameters in statistical models.

marili
Messages
3
Reaction score
0
i can perfom the same inversion in two ways:

1. using N data and k model parameters

2. using 3*N data and k+3 model parameters

where N>>k.
how can I choose the best result between the two?

thanks
 
Physics news on Phys.org
This is difficult to answer. There are some techniques which allow you to find out, statistically, whether adding another parameter does anything more than "fitting the noise" - I don't have any reference handy (~20 years ago, I did a master thesis on a very related subject but that's long ago...), sorry.

However, you can also use your common sense. If you have way enough data, it should somehow be clear whether or not you get a big improvement in the fit when you add an extra parameter or not. Maybe some intuition about the model will tell you that too.
It is only when you are a bit low on data, or with noisy data, and not sure about your model (like in black box techniques), that these issues come up.
My former mechanics professor used to say: give me 12 parameters, and I give you an elephant. Give me one or two more, and I make his thrump sling about.
 
thank you. I'm looking for some statistical parameter.
 

Similar threads

  • · Replies 2 ·
Replies
2
Views
2K
  • · Replies 13 ·
Replies
13
Views
2K
  • · Replies 6 ·
Replies
6
Views
2K
  • · Replies 2 ·
Replies
2
Views
3K
  • · Replies 6 ·
Replies
6
Views
2K
Replies
6
Views
1K
  • · Replies 6 ·
Replies
6
Views
2K
  • · Replies 2 ·
Replies
2
Views
1K
  • · Replies 6 ·
Replies
6
Views
2K
  • · Replies 29 ·
Replies
29
Views
6K