Switch from MLR to PLS Log Out | Topics | Search
Moderators | Register | Edit Profile

NIR Discussion Forum » Bruce Campbell's List » Chemometrics » Switch from MLR to PLS « Previous Next »

Author Message
Top of pagePrevious messageNext messageBottom of page Link to this message

Lois Weyer (lois_weyer)
Senior Member
Username: lois_weyer

Post Number: 42
Registered: 7-2002
Posted on Monday, May 02, 2011 - 1:24 pm:   

Oops. I just found the way to add a linear wavelength again (I have not used the system in a long time). One has to decide on the gap and segment of the first wavelength and then go to Standard Regression instead of Norris Regression. Then you can add linear terms.
And Barry, I just realized that you meant that PLS provides diagnostics for prediction outliers, not just calibration outliers. We had to do that in the 1980s's via separate programs. It was done, however. When we first started out, our online instrument would say that an empty conveyor belt had 72% nitrocellulose on it, if we didn't have a spectral matching system.
Top of pagePrevious messageNext messageBottom of page Link to this message

Lois Weyer (lois_weyer)
Senior Member
Username: lois_weyer

Post Number: 41
Registered: 7-2002
Posted on Monday, May 02, 2011 - 1:09 pm:   

Barry: I have emailed you two of my articles from the mid 1980's which mention the application of the derivative/MLR approach. I think I also have Karl's two other articles. If you can't get those, let me know. Dave: I'd like to mention that DuPont may have been reluctant to employ full-scan online instruments because they sold filter ones. At Hercules and other companies, we installed full-scan ones in the mid 1980's - both reflection and fiber-optic long-distance ones.
Hee is a summary of what the NSAS DOS "Norris Regression Program" does:
Enter file by choosing from a directory
1) Change parameters
2) Select wavelength
Under Change Parameters, you choose the gap and segment ranges. For example, a segment of 10, ranging gap from zero to 40 in ten steps. When you press "2" to select wavelength, in less than a second you get a chart showing the k(0), k (1), wavelength, R, and Std error of the wavelength with the best R. In other words, it chose for you. It was not that hard or complicated.
Then you can:
1) Continue (selects a second wavelength)
2) Plot R vs wavelength
3) List residuals
4) Plot residuals
5) Plot predicted vs lab
6) Plot control charts
7) Summary of results
8) List R for all wavelenths
9) Save R
10) Save equation

I thought that it allowed a second linear wavelength rather than just a numerator or denominator, but the version I have now doesn't seem to have that.
As you can see, it also provides diagnostics so you can clearly see outliers.
In all of my trials, PLS did not out-perform this approach, and I found the MLR equations to be much more robust, having removed interferences and emphasized the main ingredient.
If anyone wants to see this, find someone with an old NIRSystems instrument. I have one, but I am in Maryland.
Top of pagePrevious messageNext messageBottom of page Link to this message

Barry M. Wise (bmw)
Junior Member
Username: bmw

Post Number: 6
Registered: 2-2011
Posted on Sunday, May 01, 2011 - 11:47 am:   

On Apr 30, 2011, at 7:25 AM, Karl Norris wrote:

My name has received complimentary reference many times in this discussion. My good friend, Lois Weyer, has stated the advantages of derivative ratios for calibrations on diffuse reflection and diffuse transmission spectra, but here are three references:

Many thanks for the references. I'm sure I can get the JNIRS article on-line, but I'm guessing the other two are harder to locate. Do you suppose you could send a copy to [email protected]?

I have a computer program called "MULR" which performs the operations needed to do a calibration. I gave this program to Camo about two years ago, with the idea that they would include a version of this program in their "Unscrambler". I thought it would be included in their next revision at that time. Unfortunately, it has not been included, and I have not had any recent communication with Camo. I will supply a copy of this program to anyone interested. It has a steep learning curve, and operates on a computer with Microsoft operating systems. I have given this program to several researchers, but I don't think any of them are using it.

If you or Lois or Howard could write a concise description of what it does and send it to me, that would be great. From what I gather of how it works, using PLS_Toolbox as a starting point, I'd bet it could be reproduced in a couple hours given a good specification. (Reverse engineering a piece of Windows software would be much harder, especially since I don't own a Windows PC!) Coming up with a usable interface would take longer, but I'm thinking the numerical end of this is pretty simple. If somebody from this group would be willing to test it for us, we could probably have it out there in less than six months in PLS_Toolbox and Solo.

If we did that, would you buy it?

On Apr 30, 2011, at 12:05 PM, Howard Mark wrote:

Getting historical again, when PLS appeared on the NIR scene, instrument companies jumped on it and rushed to get that calibration algorithm into their software for three main reasons:

1) Gave better results (arguable, claimed by Harald Martens but at that time I didn't see any evidence of that, and from theoretical principles as I understood them then, saw no reason to expect that)


I'd be willing to bet that it DID give better results for people that weren't experts in NIR spectroscopy. And that's the key, because that's most of the audience. You can't totally supplant years of experience with statistical tools, but with the diagnostic information available in PLS plus proper validation methods you can go along way to making the most out of what you do know. Add that to the ability to identify and diagnose prediction outliers, and you've come to a working solution.

2) Avoided the need to select wavelengths. This reason was legitimate, certainly at that time, when MLR was the only other method, and we've already discussed here the difficulties that raised.

Yeup. Variable selection is hard problem compared to dealing with rank deficient data via other methods.

3) PIZZAZZ

From the instrument companies' points of view, however, the only justification for spending the time and effort needed to write the code and include PLS in their software packages was the third: the PIZZAZZ factor. That's a marketing decision and technical factors carried little or no weight.


That's bogus. Users like software that provides solutions, and they prefer it if it doesn't take them a full stint in grad school to get the background to allow them to use it properly. As for marketing, in the first 10 years Eigenvector spent about $5k a year on marketing, and during that time built up a user base that numbered in the thousands. It wasn't marketing. It was word-of-mouth, and that only happens if the tools are USEFUL.

Best regards,

BMW
Top of pagePrevious messageNext messageBottom of page Link to this message

Howard Mark (hlmark)
Senior Member
Username: hlmark

Post Number: 420
Registered: 9-2001
Posted on Saturday, April 30, 2011 - 2:04 pm:   

This response is mainly to Lois's comments, and is probably applicable to Karl's, also. Lois jogged a vague memory from when I was at Technicon, and Neotec/Pacific Scientific/FOSS (whichever name they had at the time), the creators of the Vision software that Lois mentioned, was our sworn enemy so they hardly told ME what they were doing. But we would get feedback from the field, through or own salesmen, who would contact all users, including the competitors'.

But both companies make decisions about what to put into their software, and what to keep in the software, based on the feedback they received from their customers.

So Lois jogged a memory that at some point we heard that "the enemy" took out that feature from their software, and while we didn't really KNOW why, we could make some fairly shrewd guesses. The main reason for performing such actions was that "the users", for whom the majority at that time consisted of grain elevator operators, found the feature too complicated to use, and didn't see any compensating benefit from it.

We had a similar problem in regard to measuring and setting an instrument parameter that was essentially a baseline correction, and in that case we KNEW that the customer's complaint was that it was too complicated. Therefore, even though we in the engineering and applications departments knew that it was theoretically beneficial, it was replaced by a hard-coded value based on the "average" value for several instruments, rather than allowing the customers to measure and enter the value for their specific instrument.

I think most of us don't appreciate the large distance between those of us who work in the lab (and enter on-line discussions such as this one), and the majority of users, who have neither time, interest, inclination, or resources to learn and put into practice, the things we have learned over the years.

Even people in Pharma companies, who are researchers and inclined to learn new things and so are relatively good about these things, are limited by the amount of time they have available to spend learning. Consider that these users are very much outnumbered by majority of users who are more like the grain elevator operators we dealt with. In their view, no matter how sophisticated and necessary it might be, the instrument is a tool for getting an answer. They have no inclination or incentive to learn more about it's operation than they would, say, about a thermometer. Furthermore, depending on the field (f'rinstance, biologists, perhaps), anything that smacks of the least bit of mathematics is completely anathema to them.

But since these are the majority customers, or potential customers, of the instrument companies, any decisions about what the instruments will contain or how they will be presented, will be decided based on management perceptions of what will influence these groups, not what constitutes the best science or even what necessarily gives the best results (even supposing that we could agree among ourselves what constitutes "best").

Getting historical again, when PLS appeared on the NIR scene, instrument companies jumped on it and rushed to get that calibration algorithm into their software for three main reasons:

1) Gave better results (arguable, claimed by Harald Martens but at that time I didn't see any evidence of that, and from theoretical principles as I understood them then, saw no reason to expect that)

2) Avoided the need to select wavelengths. This reason was legitimate, certainly at that time, when MLR was the only other method, and we've already discussed here the difficulties that raised.

3) PIZZAZZ


From the instrument companies' points of view, however, the only justification for spending the time and effort needed to write the code and include PLS in their software packages was the third: the PIZZAZZ factor. That's a marketing decision and technical factors carried little or no weight. The other two reasons were very far distant runners-up, although of course those were the reasons advertised and promoted. But the fact that PLS was new and different, and gave the software something other than the "tired old" MLR, regardless of merit, was all the reason that was needed - "necessary and sufficient" as they say.


\o/
/_\
Top of pagePrevious messageNext messageBottom of page Link to this message

Karl Norris (knnirs)
Senior Member
Username: knnirs

Post Number: 49
Registered: 8-2009
Posted on Saturday, April 30, 2011 - 9:25 am:   

My name has received complimentary reference many times in this discussion. My good friend, Lois Weyer, has stated the advantages of derivative ratios for calibrations on diffuse reflection and diffuse transmission spectra, but here are three references:
Spectrophotometry of human hemoglobin in the near infrared region from 1000 to 2500 nm. J.T. Kuenstner, and K.H. Norris. J. Near Infrared Spec. 2, 59-65. 1994.

Rapid measurement of analytes in whole blood with NIR transmittance. K.H. Norris, and J.T. Kuenstner. In: Leaping Ahead with Near Infrared Spectroscopy. (ed. G.D.Batten,P.C. Flinn, L.A. Welsh, and A.B. Blakeney), Published by the NIR Spectroscopy Group, 1/21 Vale Street, Melbourne,Victoria, Australia. 431-436. 1995.

Applying Norris Derivatives. Understanding and correcting the factors which affect diffuse transmittance spectra,
Karl H. Norris, NIR news 12(3), 6-9 (2001)

I have a computer program called "MULR" which performs the operations needed to do a calibration. I gave this program to Camo about two years ago, with the idea that they would include a version of this program in their "Unscrambler". I thought it would be included in their next revision at that time. Unfortunately, it has not been included, and I have not had any recent communication with Camo.
I will supply a copy of this program to anyone interested. It has a steep learning curve, and
operates on a computer with Microsoft operating systems. I have given this program to several researchers, but I don't think any of them are using it.
Top of pagePrevious messageNext messageBottom of page Link to this message

Lois Weyer (lois_weyer)
Senior Member
Username: lois_weyer

Post Number: 40
Registered: 7-2002
Posted on Saturday, April 30, 2011 - 7:04 am:   

Yes, Vision and now other software also allow one to choose a one gap and segment ar a time. What is missing is something many people probably don't know was there in the DOS version - the automated routine of going through a series of gaps at one segment and providing the user with the best choice plus a chart showing what was happening at all wavelengths. One would start at a gap of 5 nm, for example, and it would go through segments of 5, 10, 15, 20, 30, 40 or whatever (I don't remember exactly). Then you would start with a gap of 10 and do the same. Then you could choose a second wavelength, which often helped improve the model by accounting for another variable. It was very powerful and it did not require an expert to do it. One could look at which wavelengths and conditions were chosen then and see if the chemistry made sense. I don't disagree that the spectra need to be looked at, I'm saying that there was a good alternative to PLS with its own special power, and it is not there anymore. It would be very tedious to do this with the current software and I doubt that anyone does.
Top of pagePrevious messageNext messageBottom of page Link to this message

Jos� Ram�n Cuesta (jrcuesta)
Member
Username: jrcuesta

Post Number: 11
Registered: 11-2009
Posted on Saturday, April 30, 2011 - 2:39 am:   

Vision Software (NIR Systems) has for MLR the option to select segment and gaps, wavelenghts,.. at your choice. Also you can add the terms by addition or like a divisor and some other tools to make nice MLR calibrations. The problem is like Howard said that it is necessary the knowledge to add those terms, to look carefully at the spectra, to have in mind spectra of pure constituents,....,etc, all those things specialy Karl Norris and most of the good spectroscopist who writte in this Forum have.

Best Regards

Jos� Ram�n
Top of pagePrevious messageNext messageBottom of page Link to this message

Lois Weyer (lois_weyer)
Senior Member
Username: lois_weyer

Post Number: 39
Registered: 7-2002
Posted on Friday, April 29, 2011 - 7:25 pm:   

Hello everyone. It is nice to see one of my favorite topics being addressed again. You all have made some excellent points, but I'd like to add few thoughts.
I think that the combination of being able to apply a series of different gap and segment derivative treatments along with wavelength selection (and divisor terms)gave a lot more power to early calibrations than most people realize. The equations were quickly optimized using the early NIRSystems software because one could go through a number of variations quickly. That power was lost when NIRSystems went to Windows, and it never was there for other manufacturers.
Some of the power of the derivative/MLR approach was that measuring wavelengths could be found that maximized contributions from the constituent of interest while minimizing interferences. One way this worked was that sharp peaks could be accentuated at the expense of broader ones (polypropylene vs cellulose for example). Another way it worked was to create zero crossover points where the constituent of interest had a good absorption but the major interferences were crossing over the zero (node)in their derivatives. It did not take "experience" to create these situations, only running through the derivative options and taking the top choices (and then looking at what they were doing afterwards).
I'm thinking that maybe this got lost because the software was probably proprietary for one manufacturer, and then that company dropped it in favor of PLS. PLS does consider all of the wavelengths, but I think that there is still something to be said for the derivative/MLR approach in which some information is removed and some accentuated.
Top of pagePrevious messageNext messageBottom of page Link to this message

David Russell (russell)
Senior Member
Username: russell

Post Number: 55
Registered: 2-2001
Posted on Friday, April 29, 2011 - 2:07 pm:   

Let me contribute a historical perspective that has not been (I don't believe) mentioned as yet.

When I first got into Process Spectroscopy in the late 1980's, full spectrum instruments were still largely in the lab.

So folks had various favorite methods for deciding what wavelengths to choose for the interference filters that would go into a process ready filter photometer.

When we got our first taste of PLS from CPAC, no one cared that I could fit the data better with PLS since "that instrument" wasn't going out on the line.

So you ran a series of spectra and hoped that you could see a "measure" wavelength plus one or two that would serve as good references.

Some of Karl's work was implemented in the NIRSystems software (NSAS).

Howard's contributions were reflected in the Technicon software (IDAS).

We used both approaches to pick wavelengths.

We also independently developed a method where we applied simulated filters to the lab data to test whether a particular commercially available filter would generate a good result.

It was not until a few years later ~ 1993 ? that on-line NIR was good enough to apply on-line.

But we have always been careful to "validate" our models to assure that we're really measuring the chemical of interest and not an accidental correlation.

But getting back to the question of "when to swtich", we switched when we could (because on-line scanning instruments got better) or because we had to (when 5 filters just weren't enough).

But either approach requires you to know your process and respond to upsets as they occur.

MLR will let you down when a new interferent has a significant effect on one of your wavelenghts.

Similarly if the interferent absorbs strongly in a previously quiet region of your spectrum a PLS model that includes that region would also be adversely effected.

So you have to be prepared to re-model going forward and depending on the difficulty, your approach may need to change.

Better instruments and better software have made initial implementation easier, but there's still no substitute for hard earned experience.
Top of pagePrevious messageNext messageBottom of page Link to this message

Howard Mark (hlmark)
Senior Member
Username: hlmark

Post Number: 419
Registered: 9-2001
Posted on Friday, April 29, 2011 - 11:41 am:   

Barry - that's not entirely true (your comments about Karl being the only person who can use "division math"), but it's close enough to the truth, for the reasons I outlined in the previous message that you pointed us back to (thanks for that, BTW).

Karl developed his software very early on. When I started out in NIR at Technicon (in 1976), Karl already had an early version of what he ultimately came up with and that he uses now. It has been stable for, I'd say, at least 20 years at this point. It's not really very complicated; in principle, his software does MLR calculations, with options for averaging data at a few neighboring wavelengths, performing 1st or 2nd "derivative" calculations on the (averaged) data as 1st or 2nd differences on wavelength points that may be separated be several intervening wavelengths (what's called "Segment" and "Gap" in other software packages), and finally, the ability to divide a quantity at one wavelength by a quantity at another wavelength. Nothing very msyterious.

But the big difference between Karl's program and other packages is the ability, and Karl's willingness to take the time to ride close herd on what the computer is doing, and nudge it in directions that he thinks might be good directions to go in; combining letting the computer do what it's good at (computation) with letting people (Karl, specifically) doing what people are good at (thinking). Most packages kinda cut out, or at least shortchange, the second part. And most people using those packages are spoiled and WANT the computer to do their thinking for them - it's SO much easier.

Doing this successfully, of course, depends to a large extent on Karl's experience and insight into how to determine the "good" directions. But Karl wasn't born knowing that, either. He learned it by doing it lots and lots of times, and spending lots and lots of time doing it, something that most of us don't have the luxury of the time to do it, or the patience. But that's the only "magic" that Karl has. I'd say that the software contributes only 5-10% of what Karl can do.

As an employee of the USDA (i.e., US government) Karl was always willing to share his source code. As far as I know, he may still be willing to - you'd have to ask him. But that's still not going to give you his "magic".

\o/
/_\
Top of pagePrevious messageNext messageBottom of page Link to this message

Barry M. Wise (bmw)
New member
Username: bmw

Post Number: 5
Registered: 2-2011
Posted on Friday, April 29, 2011 - 11:09 am:   

On Apr 28, 2011, at 2:26 PM, Tony Davies wrote:

To Fernando my reply is that I agree with Don that there is no good time to switch to PLS! If you have a good MLR program (and I agree with Andrew that it would include division maths)

So a good MLR program would include variable selection and division maths.

and years of experience then you do not need PLS.

And there's the rub, right? If you've got years of experience doing almost the same problem over and over, then it is likely you've learned exactly where the critical information can be found in the spectra and what non-linear transform to use on it.

The problem with this is that new-comers do not have years of experience (which is difficult to teach) and they probably do not have access to the best MLR programs.

Maybe you can post a list of the best MLR programs and where to get them. We're not at all opposed to adding new tools to our software if users find them useful, so maybe we could learn something here. But we have thousands of users, and nobody has asked for this.

Although I am pleased to say that Harald Martens an I have been good friends for many years, I regret the success of PLS. I much preferred PCR which I think is much easier to understand and is more under user control.

Guess I don't know what you mean by that. PLS and PCR both have exactly the same amount of user control: you have to set the number of factors in the model. The same tools for selecting that are available in both, as are the tools for preprocessing the data, selecting variables, etc.

Again it did not have the software support which was given to PLS.

I can't speak for other software companies, but in PLS_Toolbox and Solo PCR is on the list of available methods right next to PLS and has exactly the same support.

PLS may give slightly better results but is this advantage worth the extra complexity?

This statement is wrong on two counts. First, PLS does not give slightly better results than PCR. In fact, you can't show that the prediction errors are on average lower for PLS than for PCR. I tried showing this many years ago, and it didn't work out. This has been confirmed by others. See for example, Peter D. Wentzell and Lorenzo Vega Montotl, "Comparison of PCR and PLS Regression through generic simulations of complex mixtures", ChemoLab, Vol. 65, pps. 257-279, 2003. Second, maybe the math behind the algorithms is somewhat more complex, but PLS models are typically less complex than PCR models in that they have fewer factors, and generally no totally extraneous factors. PLS generally gets the nod over PCR because it is easier to use and more computationally efficient. The choice of number of factors is usually more obvious, and the relevant model information is concentrated in fewer factors.

The annoying fact is that PLS was really powered by the take-up of NIR analysis by the pharmaceutical industry

In my experience it was the petrochemical companies that saw a big advantage in PLS with applications like gasoline octane. This was going on in mid-1980s. At that time, I was involved with CPAC, the Center for Process Analytical Chemistry, at the University of Washington. The sponsors saw the PLS program available through CPAC, written by Dave Veltkamp, as one of the main benefits of joining. Most of the companies involved were chemical, petrochemical and consumer product manufacturers. PLS didn't gain traction in Pharma until years later.

when MLR

plus variable selection, non-linear transforms and years of experience

should have been all that they needed.

Yeup.

BMW
Top of pagePrevious messageNext messageBottom of page Link to this message

Barry M. Wise (bmw)
New member
Username: bmw

Post Number: 4
Registered: 2-2011
Posted on Friday, April 29, 2011 - 10:14 am:   

On Apr 28, 2011, at 12:18 PM, Andrew McGlone wrote:

Don't forget about division math or similar non-linear 'MLR' type frameworks.

I had never heard of "division math" until now. Apparently this is something very peculiar to this NIR group as a Google Search turns up mostly links to this list! Do you have some published references on "division math?" From the discussion about it a year ago

http://www.impublications.com/discus/messages/5/6308.html?1268921323

it sounds like Karl Norris created software for it that only he could use? And apparently people use PLS mostly because us PLS guys are better at marketing! I'm learning a lot on this list! Guess I should have subscribed earlier....

They are often worth a look before giving up and bolting for the PLS door. Because they can create models in a non-linear space of the variables that PLS models

or any other inverse least squares model that is linear in the parameters

simply can't enter. I fear they are often mistakenly lumped under the banner of MLR, and so overlooked, since they necessarily operate on individual wavelengths selections in the same way as standard MLR does.

And another thing I've come to realize is that on this list, "MLR" actually means "variable selection plus MLR."

And I think the concern expressed in the thread, that PLS can't set the weight on a variable to zero, is perhaps overstated today. Jack-knifing schemes that cycle variables, or more usually intervals, in and out of a PLS model are very common now. Barry Wise's software

I assume you mean PLS_Toolbox and Solo

calls it iPLS, a name that is quite appropriate (i=interval) but greatly annoys me since it remains me that I don't own shares in
Apple!


Lars N�rgaard is the creator of iPLS. But, yes, it is a very nice variable selection method that really gives you some insight about where the predictive power of your model is coming from.

BMW
Top of pagePrevious messageNext messageBottom of page Link to this message

Tony Davies (td)
Moderator
Username: td

Post Number: 255
Registered: 1-2001
Posted on Thursday, April 28, 2011 - 4:26 pm:   

Hi,

I feel I ought to make some contribution to this debate. I�ve been following its progress but have been too busy to comment (my life for the last year has been taken over by: 1) the attempt to move house, 2) moving house 3) decorating the newer house and 4) tackling the over-grown garden).

To Fernando my reply is that I agree with Don that there is no good time to switch to PLS! If you have a good MLR program (and I agree with Andrew that it would include division maths) and years of experience then you do not need PLS. Karl is the acknowledged expert and he has never been beaten by PLS in an IDRC �Software Shootout�. The problem with this is that new-comers do not have years of experience (which is difficult to teach) and they probably do not have access to the best MLR programs.
Although I am pleased to say that Harald Martens an I have been good friends for many years, I regret the success of PLS. I much preferred PCR which I think is much easier to understand and is more under user control. Again it did not have the software support which was given to PLS. PLS may give slightly better results but is this advantage worth the extra complexity? The annoying fact is that PLS was really powered by the take-up of NIR analysis by the pharmaceutical industry when MLR should have been all that they needed.

I am hoping they this year I will have the time and stamina to complete our long promised paper on the application of CARNAC which we believe is the answer to the problems of the application of NIR spectroscopy by people who are not spectroscopists or chemometricians.

Best wishes,

Tony
Top of pagePrevious messageNext messageBottom of page Link to this message

Andrew McGlone (mcglone)
Advanced Member
Username: mcglone

Post Number: 22
Registered: 2-2001
Posted on Thursday, April 28, 2011 - 2:18 pm:   

Just catching up on this thread.

Don't forget about division math or similar non-linear 'MLR' type frameworks. They are often worth a look before giving up and bolting for the PLS door. Because they can create models in a non-linear space of the variables that PLS models simply can't enter. I fear they are often mistakenly lumped under the banner of MLR, and so overlooked, since they necessarily operate on individual wavelengths selections in the same way as standard MLR does.

And I think the concern expressed in the thread, that PLS can't set the weight on a variable to zero, is perhaps overstated today. Jack-knifing schemes that cycle variables, or more usually intervals, in and out of a PLS model are very common now. Barry Wise's software calls it iPLS, a name that is quite appropriate (i=interval) but greatly annoys me since it remains me that I don't own shares in Apple!
Top of pagePrevious messageNext messageBottom of page Link to this message

venkatarman (venkynir)
Senior Member
Username: venkynir

Post Number: 133
Registered: 3-2004
Posted on Tuesday, April 26, 2011 - 3:36 am:   

Dear Howard Mark;
I have gone through your Post .It is touching one.
Dedicated people should think what to be done 1. Developed country ( standard is going done ) 2. Developing Country 3.Under developed country .
What CNIRS and ICNIRS doing for them ?
How to bring them into clench NIRS .
How to convert the experimental things to Application oriented one.
I am inspired seeing Karl , Daham , You and David in CNIRS. That motivated work more in NIRS application .That should multiplied and young people should enter in .Experience and Export knowledge blend with young we can see the result and the giant will roar all over the world .
I have seen more marketing talks in NIRS than....
Top of pagePrevious messageNext messageBottom of page Link to this message

Howard Mark (hlmark)
Senior Member
Username: hlmark

Post Number: 417
Registered: 9-2001
Posted on Monday, April 25, 2011 - 12:46 pm:   

Peter - good point. But the US is what I see most of, and where most of my interest is, although the majority of short courses we did under the auspices of CfPA were were presented in Amsterdam; so they were definitely international in scope, yet the attendance eventually dwindled away there, too.

I agree that a short course given on the Internet may well be the kind of new approach we need, although I've never heard of Moodle. One of the problems that I perceive as limiting the penetration of educational efforts comes under the heading of "Marketing". We technical types tend to ignore or even worse, look down on, marketing but basically it consists of the ways you find new people who may be interested in something you've got, plus ways to present it to them. Put in those terms it becomes obvious why commercial companies spend so much time and effort at it, but it also makes it clear that it's what we're doing when we try to "sell" the concept of education to potential NIR users.

Having a course on the internet may indeed make it more attractive, but even the best course will not attract students if they don't know about it. Getting that message across is "marketing".

Maybe this initiative is something that CNIRS and ICNIRS can do cooperatively, too, although it sounds like CNIRS would be getting in a bit late in the game. Still, late is better than never. Maybe you can contact Gabor Kemeny and see if you can drum up some interest; Gabor is current Chair of CNIRS and you seem to know more about the Moodle effort than anyone else.

\o/
/_\
Top of pagePrevious messageNext messageBottom of page Link to this message

Peter Flinn (peterf)
Junior Member
Username: peterf

Post Number: 7
Registered: 6-2004
Posted on Monday, April 25, 2011 - 5:15 am:   

As it seems a few "grumpy old men" are waxing lyrical on the issue of NIR education, I thought I should throw my hat into the ring - despite in NIR terms still a youngster compared to some!

Howard - I agree with a lot of your frustrations, but you have concentrated entirely on what has happened (or not happened) in the good ol' US of A. May I gently remind you that the world consists of other countries as well? You mention past (and laudable) efforts of CNIRS. There is also ICNIRS!

The need for improved education of NIR users (and not just new ones!) has been recognised as a major issue by ICNIRS for years, and I admit progress in the field across the world has been glacial, despite efforts by several people over the last 8-10 years.

The problem was particularly obvious at NIR-2009 in Bangkok, and prompted newly elected ICNIRS Chairman Pierre Dardenne to use his closing speech as a very useful lecture on the basics of NIR calibration (since reproduced in NIR News, Vol 21, No 1, Jan-Feb 2010).

However, there are glimmerings of light on the horizon. The Education Sub-Committee of ICNIRS, chaired by Ana Garrido-Varo and ably supported by that "Educator Extraordinaire" Ian Murray, has been working quietly but steadily behind the scenes on this topic. As will be seen in the March edition of NIR News (Vol 22, No 2, p.19) an international virtual platform for teaching and learning NIR in "Moodle" is being developed jointly by ICNIRS and the University of Cordoba, Spain. This will be demonstrated at the forthcoming NIR-2011 Conference in Cape Town, South Africa (another reason to attend!)

I am not saying this will solve all the problems of NIR education overnight, but an internet-based programme as an alternative to traditional short courses will certainly be more accessible, and MAY be more successful. Obviously, this will depend on the content and how it is presented, and is only in its early stages.

However, in good old Aussie parlance, let's "give it a go"! It just might work.

Cheers, Peter (Immediate Past Chairman, ICNIRS)
Top of pagePrevious messageNext messageBottom of page Link to this message

Howard Mark (hlmark)
Senior Member
Username: hlmark

Post Number: 416
Registered: 9-2001
Posted on Sunday, April 24, 2011 - 12:19 pm:   

Don - The problem of educating new users in NIR started with the inception of NIR and continues to this day. Partly it�s due to the fact that NIR didn�t start out in academia, as most other analytical techniques did. We all know that Karl Norris was working in the USDA when he came up with the �magic� combination of NIR spectroscopy, reflection measurements and (what we now call) chemometric data analysis, to launch the methodology.

This sidestepped the usual development of the technique in academic laboratories, so that when it became an industrial and commercial success, there was little or no supporting academic constituency in place to help newcomers learn about it. As far as I know, there still isn't.

Attempts were made to correct that situation. When the NIR Council (as CNIRS was originally called) formed, it attempted to develop various projects to promote progress in and disseminate information about NIR. One of those was Dave Wetzel�s attempt, as an academician himself, to interest other professors and chemistry departments in looking at this new technique; he developed a set of teaching materials and made them available to anyone associated with an academic institution, who would make them part of an advanced analysis course. The response to this initiative was a resounding silence. I don�t think he even got one person to express any interest.

Over the years, several of us got involved in other attempts. Almost simultaneously with Dave Wetzel�s initiative, a short course in NIR was offered under the auspices of ACS, with various of the �old-timers� (Don Burns, Jerry Workman, Emil Ciurczak, myself and several others) rotating through as instructors. This course, and other publicly offered short courses (at FACSS, EAS, etc) struggled to get students. Arguably the most successful one was organized by Ed Brame and offered under the auspices of the Center for Professional Advancement (CfPA). I lost track of exactly when it started, but at various times included myself, Lois Weyer and Emil Ciurczak as co-instructors for about 6-7 years, maybe even ten. After Ed�s passing, Emil & I kept it going for another 5-6 years, until decreasing enrollments caused CfPA to cancel it. All these courses included a mix of spectroscopic theory, chemometrics, applications and miscellaneous other topics.

The point is that unlike courses in Chromatography, FTIR, and other types of analysis, which seem to enjoy overflow attendance, these NIR courses all (at best) struggled, and there were no other available venues.

Some instrument manufacturers provided courses in the use of their own instruments, at least in the early days. I don�t know if any of them do even that much now. Taking the 1-week course that Technicon offered while I was employed there as typical, most of the course covered how to push the buttons, how to tell when the instrument was working correctly, and how to grind samples. Those practical aspects were good, but the only exposure the class had to the spectroscopy, calibration concepts and even the chemometrics was a single 3-hour lecture I gave around Thursday mornings.

I have no reason to think that the situation is any better now.

So I agree with your assessment of the state of NIR knowledge and education. The CNIRS is revisiting the idea of sponsoring a short course, but unless we can identify the reasons people don�t flock to such courses despite the clear need, I don�t know if they will have any better success than has been achieved over the years.

I can�t think of any major new or different ways to try. I suspect that a large part of the problem is the perception on the part of would-be NIR users that except for the chemometrics, none of the other topics are important. This isn�t entirely their own fault. Especially in the early days, and probably still continuing to this day, the message from the instrument companies as well as the NIR community, and being reinforced by the chemometrics community, is that those topics aren�t important, because whatever faults or failures might occur in any of those other arenas can be covered up and compensated for with the mathematics. For better or worse, there�s a fair amount of truth to that, enough to convince the newbies that the algorithms contain all the �magic answers� to their problems.

So the newcomers have little incentive or justification to learn about anything beyond the boundaries of Chemometrics, and established scientists also seem to have little or no interest or incentive to do research into any fundamentals of the spectroscopy. Not like it was, say, 50-60 years ago when mid-IR spectroscopy was new. Lots of papers on new and improved whiz-bang algorithms, though. Our recent paper (for those who may have missed it, see Appl. Spect.; 64(9); p.995-1006 (2010)) is the only one in recent memory that addressed any fundamental aspects of the spectroscopy of NIR or its relation to other disciplines. Even JNIRS contains applications papers almost exclusively. Ian Michaels laments this, too, but we�ve spoken about it and he informs me that that�s what�s being submitted.

Maybe you can think of something different to try, from your point of view.

\o/
/_\
Top of pagePrevious messageNext messageBottom of page Link to this message

Barry M. Wise (bmw)
New member
Username: bmw

Post Number: 3
Registered: 2-2011
Posted on Friday, April 22, 2011 - 4:22 pm:   

Hello Donald and Everybody:

On Apr 22, 2011, at 8:11 AM, Donald J Dahm wrote:

>However, as for my old man grumpiness, I see an overwhelming majority
>of new NIR practitioners, who as they enter the field, learn about the
>number crunching aspects of the chemometric techniques, and pay very
>little attention to the spectroscopy.

Yeah, I can believe that. I suffer from that myself, but I'm not really a spectroscopist
to begin with. I'm a chemical engineer and sort of a closet analytical chemist. But we
do try to understand WHY calibrations work (or not). Often that means trying to educate
ourselves about the particular system at hand and/or working with the client who has a
better understanding of the chemistry. And of course we do have 3 people at Eigenvector
who actually ARE spectroscopists, so I've always got access to somebody whose brain
I can pick!

>The old timers seem to tend
>toward MLR when they introduce spectroscopic information into their
>calibrations. Speaking for myself, when doing PLS, I have to have a
>reason for excluding data (like absorbance being near the limiting
>value) or I leave it in. When I do MLR, the opposite is true: I only
>include data if there is a reason to include it. Maybe that�s not
>good, but it�s the way I do it.

Interesting. How did you arrive at that methodology? PLS is somewhat more robust
than MLR, it has some ability to ignore bad data provide that you don't keep too many
factors in the model. It is an interesting exercise to see how an outlier sample's
influence on the model (leverage) increases as you add factors to the model. (In fact,
we do this as an exercise in our courses.) But in general I'd say it is at least nearly
as important in PLS to eliminate bad data as it is in MLR.

>We�re not such a new field anymore, and it seems to me that we have
>not done an outstanding job of developing an academic approach that
>combines the spectroscopy, the chemometrics, and the scattering
>theory. Maybe my grumpiness really stems from the fact that I haven�t
>done my job very well. Do you guys teach a course on the synthesis of
>Chemometrics and Spectroscopy in Calibration? Who does?

We of course use a lot of examples from spectroscopy in our courses and try to illustrate
how various "non-dealities" manifest themselves and how the methods and various
preprocessing techniques deal with them. Probably the closest we get to what you
describe is our Advanced Preprocess class.

At EigenU we really try to help students develop a solid understanding of the
chemometric methods they are using. This is important in interpreting the models
and understanding why/how/when things are likely to go wrong. To some extent
we assume that the students will have a good understanding of their data, and if
they add a good understanding of the chemometric methods they'll be in a good
position to learn something about the system they are working with. (In fact we
discuss this in our introductory material.)

So here's a scary thought: there are people out there doing spectroscopic
calibrations who understand NEITHER the spectroscopy NOR the chemometrics!

Happy Easter!

BMW

http://eigenvector.com/courses/EigenU.html
Top of pagePrevious messageNext messageBottom of page Link to this message

Donald J Dahm (djdahm)
Senior Member
Username: djdahm

Post Number: 58
Registered: 2-2007
Posted on Friday, April 22, 2011 - 10:10 am:   

Barry:
I have no desire to denigrate Chemometrics. If it were not for things like MLR, PLS, etc, there would be no interest in the things I do related to scattering. I have great respect for you and your discipline. Furthermore, I agree that we can use Chemometrics to extract spectroscopic information from our data, and though this gain a better understanding of our system.

However, as for my old man grumpiness, I see an overwhelming majority of new NIR practitioners, who as they enter the field, learn about the number crunching aspects of the chemometric techniques, and pay very little attention to the spectroscopy. The old timers seem to tend toward MLR when they introduce spectroscopic information into their calibrations. Speaking for myself, when doing PLS, I have to have a reason for excluding data (like absorbance being near the limiting value) or I leave it in. When I do MLR, the opposite is true: I only include data if there is a reason to include it. Maybe that�s not good, but it�s the way I do it.

We�re not such a new field anymore, and it seems to me that we have not done an outstanding job of developing an academic approach that combines the spectroscopy, the chemometrics, and the scattering theory. Maybe my grumpiness really stems from the fact that I haven�t done my job very well. Do you guys teach a course on the synthesis of Chemometrics and Spectroscopy in Calibration? Who does?
Don Dahm
Top of pagePrevious messageNext messageBottom of page Link to this message

Ana Garrido-Varo (garrido)
Junior Member
Username: garrido

Post Number: 8
Registered: 1-2006
Posted on Friday, April 22, 2011 - 4:07 am:   

Ian: I assume that at the end of my message I did not clik "post message" but...maybe your intelligent server keep the previewed post.... We will see..

Ana
Top of pagePrevious messageNext messageBottom of page Link to this message

Ana Garrido-Varo (garrido)
Junior Member
Username: garrido

Post Number: 7
Registered: 1-2006
Posted on Friday, April 22, 2011 - 4:03 am:   

I posted a long message yesterday and it dissapeared from that forum !!. I did not save copy. (bad student behaviour!!!!).... Ian could you find it in your intelligent server?? . I will be very happy.

Regards

Ana
Top of pagePrevious messageNext messageBottom of page Link to this message

Fernando Morgado (fmorgado)
Advanced Member
Username: fmorgado

Post Number: 25
Registered: 12-2005
Posted on Thursday, April 21, 2011 - 9:09 am:   

Hello :

Thanks for all your comments.
I think this is the better discussion group. The only bad of this group is always, after ask something, I realize how little I know
Thanks
Top of pagePrevious messageNext messageBottom of page Link to this message

Howard Mark (hlmark)
Senior Member
Username: hlmark

Post Number: 415
Registered: 9-2001
Posted on Thursday, April 21, 2011 - 4:36 am:   

I have to agree with Pierre, and I think that all of the other "old-timers" would also agree. Also, as Barry pointed out, in the limit of sufficient data, the MLR and PCR models are the same. When you use less than all the data you reject some information. The two approaches (MLR vs PCR/PLS) differ in the information that is rejected, and the early work on comparisons of the two approaches indicated that the advantage is not always with the full-spectrum methods.

The major problem with MLR is the practical one: how to select the set of "best" wavelengths from a large starting set? A "brute force" approach, of course, runs into the factorial explosion of combinations that need to be tested. Efficient algorithms (e.g., simplex optimization, genetic algorithms, etc.) exist but MLR went out of favor just about when those algorithms started to appear in software implementations so they were never used in more than a few research papers.

An aspect of MLR that never gets mentioned is that once you select a good set of wavelengths, then there is NO noise contribution from any of the other wavelengths, that don't contribute to the modeling capability.

\o/
/_\
Top of pagePrevious messageNext messageBottom of page Link to this message

Pierre Dardenne (dardenne)
Senior Member
Username: dardenne

Post Number: 60
Registered: 3-2002
Posted on Thursday, April 21, 2011 - 4:01 am:   

Dear All,

I agree with Don. I have been trained in the beginning of the 80's to use MLR because PLS did not exist. We did decent models using a manual search of a good set of wavelengths based on spectroscopy. PLS can fit a lot of artifacts and noise especially when the sample set is too small. The way the CV is set is of major importance. Do make the CV with blocks of spectra independent to each other; for instance sort the spectra by date and set up CV by consecutive blocks. Mark Westerhaus showed that at the last IDRC and the RMSECV are totally different regarding the sorting.
MLR is still udeful to find the more sensitive wavelengths on feasibility studies. When using several hundreds or thousands of samples, PLS is OK and looking the Bcoefficients, one can observe that the coefficients are close to zero when there is no information in parts of the spectra.

The comments and maths presented by Barry are very interesting. But we have to assume that MLR can not be used without variable selection.

Pierre
Top of pagePrevious messageNext messageBottom of page Link to this message

Barry M. Wise (bmw)
New member
Username: bmw

Post Number: 2
Registered: 2-2011
Posted on Wednesday, April 20, 2011 - 10:52 pm:   

Dear Fernando:

Thanks for asking a great question. My response can be found at:

http://www.eigenvector.com/evriblog/?p=405

Thanks!

BMW
Top of pagePrevious messageNext messageBottom of page Link to this message

Fernando Morgado (fmorgado)
Advanced Member
Username: fmorgado

Post Number: 24
Registered: 12-2005
Posted on Wednesday, April 20, 2011 - 7:39 am:   

Dear Donald :
Thanks for your answer, is a fine answer.
Fernando
Top of pagePrevious messageNext messageBottom of page Link to this message

venkatarman (venkynir)
Senior Member
Username: venkynir

Post Number: 132
Registered: 3-2004
Posted on Tuesday, April 19, 2011 - 7:00 pm:   

Hi Fernando,

This question of MLR or PLS discussed at length earlier also.
My suggestion is take spectra do pre-processing then use SPA (successive Projection Algorithms ) for variable selection .
This will suggest contributed one then use MLR.
PLS is good but I have my doubt about prediction and Validation
Top of pagePrevious messageNext messageBottom of page Link to this message

Donald J Dahm (djdahm)
Senior Member
Username: djdahm

Post Number: 57
Registered: 2-2007
Posted on Tuesday, April 19, 2011 - 4:15 pm:   

As a grumpy old man, I say the time to swith to PLS is when you are ready to admit that you don't have the knowledge or patience to do actual spectoscopy.
Top of pagePrevious messageNext messageBottom of page Link to this message

Fernando Morgado (fmorgado)
Advanced Member
Username: fmorgado

Post Number: 23
Registered: 12-2005
Posted on Tuesday, April 19, 2011 - 3:21 pm:   

Hello David :
Thanks. The instrument have 1100 datapoints enable for make a Model. The aplication is easy, Moisture and Fat. MLR work very fine and PLS too.
The question was for to know when use PLS or MLR. Always I think , with a correct Factor Number selection , PLS is more robust than MLR.
Thanks a lot
Fernando
Top of pagePrevious messageNext messageBottom of page Link to this message

David W. Hopkins (dhopkins)
Senior Member
Username: dhopkins

Post Number: 189
Registered: 10-2002
Posted on Tuesday, April 19, 2011 - 3:09 pm:   

Hi Fernando,

I don't think there is any 'rule' for switching from MLR to PLS. It depends on whether your current MLR calibrations are delivering satisfactory performance, and may depend upon your hardware. Does your application require the limited number of terms of a MLR calibration? Do you have an instrument that gives you many wavelengths? Or just a selection?

I would put your question the other way around. When does it make sense to recalibrate using PLS rather than MLR? As soon as you have 5 or more wavelengths, you can reap the benefits of PLS or PCR calibrations, which permit the use of all the wavelengths in a robust calibration that does not require the selection of a subset of wavelengths. And, I would hasten to add, I don't recommend using all the wavelengths that your spectrometer is capable of delivering. You should not employ any wavelengths that contribute only noise to the measurement you wish to make.

Best regards,
Dave
Top of pagePrevious messageNext messageBottom of page Link to this message

julen arruabarrena (julen)
New member
Username: julen

Post Number: 5
Registered: 11-2010
Posted on Tuesday, April 19, 2011 - 3:08 pm:   

Hello!

I'm new in the NIR area, but the main advantage of PLS is the data dimension reduction, so it's suggested to use when the number variable is high and highly correlated (e.g. NIR spectra)
Top of pagePrevious messageNext messageBottom of page Link to this message

Fernando Morgado (fmorgado)
Advanced Member
Username: fmorgado

Post Number: 21
Registered: 12-2005
Posted on Tuesday, April 19, 2011 - 2:52 pm:   

Dear Sirs : I have a single question.
In wich moment you consider is necesary move from traditional MLR to PLS.
Depend of the number of samples in the database? depend of the aplication?, etc.
Thanks in advance
Fernando

Add Your Message Here
Posting is currently disabled in this topic. Contact your discussion moderator for more information.