Perhaps a different approach to model development

[davidlincoln@mac.com]

I have experimented a bit with the sensor, SCIO lab etc.

 

I studied “Hard Cheese ” and though about the model development with respect to a Designed Experiment  ( DOE ).

 

 

Using a designed approach may help “crowdsource” the data so to speak, with the main advantage here is that DOE helps establish a common language and then organizes the data for model and statistics  like F test, etc. This approach is widely used  in the  chemical industry.

 

The input scheme in SCIOLab is concise and optimizes the web portal for data, but it is a bit hard to see the full experiemtal   Space with to attributes and contributes to noise from random sampling, ( but I do understand the nature of the business plan )

 

So for example with  ” Hard Cheese” and using DOE  I would present the attributes as factors ( quantitative and qualitative ) and grouped and organized as Class Factors.

 

F1. ( qualitative ).   8 Brands

 

F2.  ( quantitative )  8 levels  of fat

 

F3.  ( quantitative ).  7 levels  of protein

 

F4.  ( quantitative ).  3 types  of milk

 

F5.   ( quantitative ).   11   Sample dates

 

F6.   ( quantitative  ).    10  sampling temperatures   ( some discussion here on possibly  being qualitative )

 

 

Now doing this , one can concentrate on factors that help understand variability with respect to NIR and increase the likelihood of success of crowd- sourced method development which I think is a major component of your business plan and one that I think is “way” cool  and extremely powerful if it is carefully managed. To confound matters  further is the hardware perspective which has already been seen with V  1.1

 

In short starting to think about the crtical factors that are at play with NIR coupled with hardware and method development will help improve success. There are very big efforts with designed experiments and companies that specialize in this .

The sample identification is extremely important but the designed experiment comes first and then a sample ID is assigned.

 

Lastly this approach lends to an organization of sampling that helps testing models for statistical significance. There are some very  bright folks who do this that could help the crowd -sourced randomization ” effect”  that could give  model development  a  run. for its money.

 

 

 

 

 

 

 

 

[Author]

Posts