- This topic has 8 replies, 5 voices, and was last updated 8 years, 1 month ago by
.
-
Posts
-
I will try to describe the initial system and then ask the question.
You have a substance that has 10 different characteristics. You have lots of data and some of your data seems to be high prediction accuracy. I think if I only use factor A data, the results of factor B can be determined, but without the elimination of factors C D and E the extra data of ACDE confuses the results when determining factor B.
Does this make sense to anyone but me??
ps: You blew my mind, my time to blow yours….
If A is detectable by SCiO in the presence of CDE, and B is correlated with A, then B should also be detectable by SCiO.
SCiO’s algorithms do not use any information other than the spectra and the reference data, if there’s correlation between the reference data and the processed spectra – the algorithms should be able to find it.
Please feel free to reach out to us at dev@consumerphysics.com if you have specific questions about your experiment and we’ll be happy to advise if we can.
Lets try the question again.
I have created two models. The first model will produce results like the sample contains %6043 . But only sometimes. The rest of the time it is pretty close.
The second model quantifies another way but..
If the first model shows %6043 than the second model is actually measuring substance B, not substance A.
So, most of the time I will be measuring substance A. When it goes WAY out of whack, I know I found substance B, not A. The second model should be used for the quantity of B, Otherwise the model results should be used for substance A.
The substance has both A and B molecules in it, but it will either be predominately A or mostly B.
If the substance is B, it hides the amount of A (and I get out of wack numbers for A) but substance A does not hide substance B.
I decided to ask it in another different way:
I have a model that quantifies the amount of salt in solution that works.
I also have a model that quantifies the amount of sugar in solution.
I have noticed the sugar model doesn’t work, but the Salt model works great.. Most of the time. Every once in a while I get a result that is just stupid. Once in a while it will say something like the solution contains %6000 salt. Obviously wrong…
The thing is, when the salt reading is obviously wrong, the sugar reading is spot on…
Only one reading is correct, either the sugar or the salt….
I need a way to show one model or the other but not both…
So you need a way for the app to identify a false positive and switch models in that case? If it contains >x% sugar, switch to the sugar model which ignores the amount of salt?
If false positives are possible, there clearly needs to be a way to handle this, although it sounds like there might not be a good answer aside from simply telling the user about it.
–Scott.
Let me just check if I’m understanding this right:
You are scanning one of 2 substances: “Salt Solution” and “Sugar Solution”.
You have 2 Models- “Model A” which measures “Salt Content” and “Model B” which measures “Sugar Content”
As you do not know which solution is which, you use Model A, which measures “Salt Content”
If the value returned is within reasonable/expected parameters, you know that the substance is “Salt Solution” and the returned value can be trusted.
If the value returned is not within reasonable/expected parameters (e.g. 6000%) you know that the substance is not “Salt Solution” and is therefore “Sugar Solution” – In other words the wrong substance being measured by a model based on “Salt solution” is the cause of the unexpected results.
In this case, you scan the same solution again with Model B to determine it’s “Sugar content”
Model B returns a reasonable value for “Sugar Content” – as the substance matches the model.
Is this correct?
From what I can make of SCiO so far- at the moment this is the only way to go about it. Any sample you scan is being scanned in relation to a specific model, as opposed to picking a model based on the results. So by scanning the whole world with your “Salt solution” model you are going to be determining the salt content in everything- but not accurately, as of course the molecular makeup of everything else is wildly different and will throw the results.
My suggestion would be to have a simple identification model first – which categorizes a sample as “Salt solution”, “Sugar solution” e.t.c. Then rescan the sample against the appropriate estimation model.
I am not sure, but it might even be possible using the SDK to develop an app which uses the original scan results against 2 models. so your app would:
Scan Substance
Load scan results against the identification model and identify
Then IF Substance type=”Salt solution” load results into “Salt solution” model
Then IF Substance type=”Sugar solution” load results into “Sugar solution” model
Display: SUBSTANCE IDENTIFIED AS SUGAR SOLUTION. *Display values from Sugar Solution model.
I will try to describe the initial system and then ask the question.
You have a substance that has 10 different characteristics. You have lots of data and some of your data seems to be high prediction accuracy. I think if I only use factor A data, the results of factor B can be determined, but without the elimination of factors C D and E the extra data of ACDE confuses the results when determining factor B.
Does this make sense to anyone but me??
ps: You blew my mind, my time to blow yours….
<br>
I have no idea, sorry…
- You must be logged in to reply to this topic.