Degree of Influence and Sensitivity Analysis

The engine.
andrew_ling
Posts: 9
Joined: Wed May 14, 2008 11:16 pm

Degree of Influence and Sensitivity Analysis

Post by andrew_ling »

Hello
I'm wondering if there is any functionality in the Smile engine for calculating degree of influence and doing sensitivity analysis.
shooltz[BayesFusion]
Site Admin
Posts: 1417
Joined: Mon Nov 26, 2007 5:51 pm

Re: Degree of Influence and Sensitivity Analysis

Post by shooltz[BayesFusion] »

andrew_ling wrote:I'm wondering if there is any functionality in the Smile engine for calculating degree of influence and doing sensitivity analysis.
Sensitivity analysis is in SMILE, but at the moment is not exposed through public interface of the library (the relevant header file is not included in the download).

The code for calculating strength of influence is part of GeNIe (not in SMILE).
andrew_ling
Posts: 9
Joined: Wed May 14, 2008 11:16 pm

Post by andrew_ling »

Are there any plans to make the sensitivity analysis functionality publicly available as part of Smile.Net?
shooltz[BayesFusion]
Site Admin
Posts: 1417
Joined: Mon Nov 26, 2007 5:51 pm

Post by shooltz[BayesFusion] »

andrew_ling wrote:Are there any plans to make the sensitivity analysis functionality publicly available as part of Smile.Net?
No, there are no such plans at the moment.
lir24
Posts: 9
Joined: Mon Feb 25, 2013 5:00 am

Sensitivity Analysis tornado in jSmile

Post by lir24 »

Hi all,
I am now using jsmile for my course project, and found there is no class or methods for sensitivity analysis. I wonder if there is any method to draw the sensitivity tornado in jsmile, just like that in GeNIe?
Thank you!
Martijn
Posts: 76
Joined: Sun May 29, 2011 12:23 am

Re: Degree of Influence and Sensitivity Analysis

Post by Martijn »

Hi,

The tornado diagram is only available in GeNIe, not in SMILE.

Best,

Martijn
lir24
Posts: 9
Joined: Mon Feb 25, 2013 5:00 am

Re: Degree of Influence and Sensitivity Analysis

Post by lir24 »

Martijn wrote:Hi,

The tornado diagram is only available in GeNIe, not in SMILE.

Best,

Martijn

Oh, I see..Thank you so much~~
arsham
Posts: 3
Joined: Fri Aug 15, 2014 8:24 am

Re: Degree of Influence and Sensitivity Analysis

Post by arsham »

lir24 wrote:
Martijn wrote:Hi,

The tornado diagram is only available in GeNIe, not in SMILE.

Best,

Martijn

Martijn, is there any document or paper that explains the Tornado graph of the GeNIe sensitivity analysis tool and the theory behind it? The documentation of GeNIe regarding the sensitivity analysis is too old and does not represent the tool's features anymore.

Great thanks in advance
Martijn
Posts: 76
Joined: Sun May 29, 2011 12:23 am

Re: Degree of Influence and Sensitivity Analysis

Post by Martijn »

Hi Arsham,

This article should give you an overview of the background of tornado diagrams
http://en.wikipedia.org/wiki/Tornado_diagram

Martijn
shooltz[BayesFusion]
Site Admin
Posts: 1417
Joined: Mon Nov 26, 2007 5:51 pm

Re: Degree of Influence and Sensitivity Analysis

Post by shooltz[BayesFusion] »

[quote="arsham"]is there any document or paper that explains the Tornado graph of the GeNIe sensitivity analysis tool and the theory behind it?[quote]

The sensitivity analysis in SMILE/GeNIe is based on section 4 from the "Making Sensitivity Analysis Computationally Efficient" paper by Kjaerulff and van der Gaag (available at http://arxiv.org/pdf/1301.3868.pdf).

The tornado chart shows the top N parameters in the network influencing the specified output (the node/outcome pair) selected from the dropdown list in the upper part of the window. Below the tornado you can change how much the parameters can change (by percent or going to full 0..1 range). Note that each parameter is assumed to be changing separately - we do not perform n-way sensitivity at this point.

Let me know if you need more specific info on the tornado/sensitivity.
arsham
Posts: 3
Joined: Fri Aug 15, 2014 8:24 am

Re: Degree of Influence and Sensitivity Analysis

Post by arsham »

Thanks Shoolts for the reference and the explanation. That helps a lot. However, I still have something to ask.

You mentioned
The tornado chart shows the top N parameters in the network influencing the specified output
, and that is what I was expecting also. However, the Tornado chart actually showing the top N "scenarios" in the network and not just parameters; and by Scenario I mean an state of a parameter that influences the target node, given observed states of all of the parent nodes of the parameter. So, how can we interpret this? What are the parameters that the target node is most sensitive to, when what we actually see is a scenario that a single parameter?

I hope I was clear on this. Otherwise kindly let me know and I will gladly attach or send a snapshot of a Tornado chart of my model to visually explain what I mean.

Great thanks in advance
shooltz[BayesFusion]
Site Admin
Posts: 1417
Joined: Mon Nov 26, 2007 5:51 pm

Re: Degree of Influence and Sensitivity Analysis

Post by shooltz[BayesFusion] »

arsham wrote:by Scenario I mean an state of a parameter that influences the target node, given observed states of all of the parent nodes of the parameter.
We don't assume all parent nodes of the node containing the changing parameter will be observed. Look at this from the low level point of view: you're changing single numeric parameter in some CPT (other parameters in the corresponding column will have to change to ensure the probabilities sum to 1.0). What is calculated by the sensitivity analysis is the spread of posterior probabilities of specified target node(s). The tornado displays the parameters which yield largest changes in the posterior.
So, how can we interpret this? What are the parameters that the target node is most sensitive to, when what we actually see is a scenario that a single parameter?
That would be the correct interpretation - just keep in mind that we assume changes in single parameter (our sensitivity implementation is not N-way sensitivity).
arsham
Posts: 3
Joined: Fri Aug 15, 2014 8:24 am

Re: Degree of Influence and Sensitivity Analysis

Post by arsham »

great.
Thanks for the explanation.
Meiner
Posts: 2
Joined: Thu Jan 19, 2017 1:57 pm

Re: Degree of Influence and Sensitivity Analysis

Post by Meiner »

Hi,
I have a further question to the output of the sensitivity analysis. Once you performed the analysis and open the tornado diagram another window opens when you hover with the mouse over any of the bars of the diagram exposing values for things like:
target value range; width; parameter range; definition index and so on. Is there anywhere an explanation to what all these values and parameter meand/indicate?
I had a look at the paper of Kjaerulf, but it is VERY technical, especially for someone like me who's not so experienced in all these mathematical terms.
I'd really appreciate some more laymen terms explanation. Thanks in advance
shooltz[BayesFusion]
Site Admin
Posts: 1417
Joined: Mon Nov 26, 2007 5:51 pm

Re: Degree of Influence and Sensitivity Analysis

Post by shooltz[BayesFusion] »

Here's the info on the values displayed in the tooltip which shows up when you hover over the tornado chart bars:

1. target value range: the min/max posterior values for the node/outcome selected in the dropdown list above the tornado. The minimum and maximum depend on the changes allowed in parameters by the selection you make using the 'Parameter spread' slider below the tornado.

2. parameter range: min/max parameter value. Again, this is controlled by parameter spread.

3. current parameter value: the probability in the CPT of the node at specified index. The index is a linear representation of the parent configuration displayed above the tornado bar (after the | sign).

4. derivative: the value of the first derivative for the function which outputs the target posterior given the parameter represented by tornado bar. This function (not the derivative) has the general form T=(a * p + b) / (c * p + d), where T is target posterior and p is the parameter under study. Sensitivity code calculates these coefficients. When they're known it's trivial to obtain the derivative, which is the basic measure of sensititivy, and target posterior range (see 1 above)

5. coefficients: a,b,c,d values for the function described above.
Post Reply