Recent Posts

Pages: [1] 2 3 ... 10

HUGIN GUI Discussion / Entering different evidence

« Last post by Holly on August 14, 2023, 23:34:40 »

Hi,

I have a set of data that represents information on dams (see attached txt). I am interested in modelling how the dam `Status_ID` and `Raise_ID` affect probabilities of failure, and the type of failure mode (`FM`) if it does fail. I use the EM learning algorithm to add experience data from the csv to all the nodes (see attached network).

I want to incorporate the fact that I know I have an unbalanced dataset (failures are overrepresented) and that a true overall failure rate (for all types of `Status_ID`, `Raise_ID` and `FM`) is 2E-4. How do I best incorporate this knowledge in this network? I've tried 1) entering it as a likelihood in 'Run' mode or 2) entering it as a prior before adding experience data.

This is an example of how I would use Bayes equation by hand:
P(Failure=Yes | Status_ID="Active" & Raise_ID="Upstream") = P(Status_ID="Active" & Raise_ID="Upstream" | Failure="Yes")*P(Failure="Yes") / P(Status_ID="Active" & Raise_ID="Upstream")
where P(Failure="Yes") = 2E-4, and other probabilities come from the data

Appreciate any help!

Thanks

General Discussion / Re: Parameter Selection and Cross-Validation

« Last post by Anders L Madsen on August 02, 2023, 16:49:55 »

Hello Holly,

We do not have a tutorial showing the process for performing feature selection for a Bayesian network.

If we assume that you are considering a classification problem, then one approach is a wrapper search. Start with a model containing only the target variable and iteratively adding the feature which improves the score the most. Stop when there is no further improvement.

You can split the data into train, validate and test sets to measure the performance.

Best regards
Anders

General Discussion / Parameter Selection and Cross-Validation

« Last post by Holly on August 01, 2023, 22:31:51 »

Hello,

Are there any tutorials or examples that outline a process for performing parameter selection for a Bayesian Network? (i.e. test different subsets of potential input variables and find the subset that gives the best model performance evaluated using something like k-fold cross validation, for example).

HUGIN GUI Discussion / Bug with control points zoom

« Last post by PaN00B on June 26, 2023, 07:54:09 »

For circular fisheye lens when adding control points it shows zoom window, in the center of image its okay, but towards the very edge of circular fisheye lens zoom window shows wrong portion than where i click for control point of the image.

Bug only appears if image is set.
Lens Type: Circular Fisheye
Focal length: 1.2
Focal length multiplier: 4.099

2022.0.0.9bb147ae7161 built by Thomas

HUGIN GUI Discussion / Crop window bug

« Last post by PaN00B on June 26, 2023, 06:45:48 »

First problem i create panorama from image, after opening image and going to Masks i see the image loaded, then i create new project without saving previous one and load new image, on the Photos tab i see the new image, however when i go to Masks tab and click on image, it still shows the preview of first image from previous project that is no longer open.

Second problem in Masks > Crop if i click anywhere on the image mask disappears and i have to select image again to show it.

2022.0.0.9bb147ae7161 built by Thomas

General Discussion / Re: Parameter Learning

« Last post by Anders L Madsen on June 17, 2023, 05:19:30 »

Hi Marco,
You can run the EM learning Wizard that has the option to reset all tables and all experience counts.
Best regards
Anders

General Discussion / Parameter Learning

« Last post by MarcoGV on May 19, 2023, 18:25:58 »

I have a rather large model built (using structure learning in Hugin) using a large dataset. I would like to use the same structural model with different data. For this particular case, I would like to ignore all parameter values learnt from the large dataset. EM learning starting from the structure with all 1s in the CPTs and zero experience counts seems to be the way to do it, but I do not want to manually replace all experience counts with zero and all entries in the CPTs with 1. Is there a way to reset CPT entries in bulk?