so first a warning you can this is my personal view or everything so you can be absolutely do not agree with anything what I'm gonna say I'm not getting offended it's absolutely normal variability has to be taken with a certain degree of a let's say variability itself second mm being very personal it's only a choice of things that can be done yesterday dr. Kline told about the fact that not everything that is possible is worth being done because in practice there is a number of different limitations that are not necessarily scientific or technical that can be also economical so just they say a view of different positions in the field now let's see if this works okay this gentleman probably you remember you recognize and his job has been mainly about variability okay because variability is a very natural phenomenon a variability is the common background to life as we heard yesterday no variability no life no evolution just Darwin could stay home Nevin had been what it is so variability in reality is the founding concept of the evolution theory life itself is variability there is no life without variability as far as we understand life today we all swim in a sea of variability I like the idea of swimming because it gives you the idea of being able to move in any direction with minimal effort and being also and gives the idea of the fact that you should try to move along with the current in a way not to have the variability work against you but to work together with variability and trying to use variability for your for the best okay never never put yourself against the current recognize that variability is there and do what what you can do to go forward so just a very tiny introduction to typical I would call it a typical but simplified feed flow so we have ingredients each ingredients has its variation in its nutrients okay then we add to that variation the sampling error we had the analytical error then we have the data we use in our optimization we take decision on this but anyway we get to a point where we have numbers that we use then we proportion the feed and we have the our rounding error in the diet then we proportion the feed and we have dosing errors scale errors imprecisions whatever you can think of then we have our mixing errors because not every mixture is the same and not every mixer is well-maintained as it should be then you have after you made the mixer and you pelletized and you've made everything you have the segregation of the nutrients due to the different parts of processing transport to farm and feeders and finally the errors of delivery that we mentioned yesterday they are not at all something that does not happen it happens it happens much more than one then we can think of so there's lots of points where we could say that there is a hot spot to this we have to add the fact that we are dealing with a with an area where that uses words in a very specific way that not necessarily is the way we use them in everyday life so we are especially when we are talking about sampling and analytics but we talk about the curacy we talk about error of measurement we could talk about random error or systematic error trueness vs precision repeatability reproducibility all these things have specific meanings in the in this contact and slightly different meanings in everyday life for example this is the typical way we've been thinking about accuracy and precision okay accuracy is just simply having the average value okay but then if it's okay you are you have a very high let me see if I can run this name yes you have a very high accuracy because the average the data is okay but you have a very poor precision because the data is catered around this reverse it's your extremely precise because you're always doing the same thing but you're doing it always wrong so this is something you don't like this is something you don't like very much at all this is your target but the bullseye is well it's difficult to attain seeing it in it from a different perspective this is what we are looking at so this is totally wrong for us this is well accurate it's not accurate but it's precise this is precise no but it's accurate yes because the average is okay and this is what we have to look for said this here we have a mix up of different causes for these scattering of data around the average there can be systematic errors and random errors you cannot correct for random errors because by definition being random you cannot but you can correct for systematic errors as far as you can find them very interesting if you can go and have read these materials from the Royal Chemical Society they are excellent I give the very excellent let's say very short and very concise things about ideas in terms of precision analytical precision but not only related to analytics this for example is a different way to think what we've just seen so it's not just a question of improving accuracy here decreasing uncertainty so reducing the error that is a systematic error built-in in the system it's difficult for us to to discriminate between all these errors within our process the process I showed you before so that's the approach we have we are we are within the system of variability and we have from within to deal with it so you can see that definitely a different approach so the problem that gets up everything that is coming from here has been taken and real aberrated from this paper here that's I think it's a seminal paper but it didn't give any follow up and it's a very very good one very helpful in my professional life anyway so everything is depending on the goals you have you're feeding the animals once and then you'll forget it and you go to different business you'll try to keep them alive you're trying to produce consistently ingredients of the certain quantity the more thing you ask the more is different the level of variability you are going to accept and they better or variability management you have to apply because what is the difference is the consistency you want your results so if you want consistent feeds we have to deal with feed ingredients variability so we have the real variability as we said then we have all the problems related with the analytics sampling etc and then the process the feed processing variability variability in feed is a complicated issue because in a variability of in certain nutrients in a feed can be calculated in this way according to Faust it's the the sum okay of a number of compounds that are the product of the standard deviation of the nutrient in the feed mixture each nutrient okay multiplied by the fraction of the total neutral contribution by the ninth ingredients that means corn has about I don't know 15 percent of the variability of the old diet so we multiply corn by the percent by its standard deviation we square it we sum everything and we put it under square root when you play with these things results becoming predictable unpredictable they are not linear you get results that can very easily go away and what is interesting is that you can make a test about it for example you see here a list of ingredients and the second column you have the typical protein content for it then you have a percentage in a formulation a very easy thing gradient formulation you have at the amount of nutrient that is delivered by each ingredient to that formulation for a total of 21 point seven percent that is considered to be a hundred this is the protein percent coming from the different ingredients out of the 100 this is the coefficient of variation ie the standard deviation in units okay.you Newton's units and then you see that we had you have this value as the sum of all these ingredients multiplied so it's the percentage of protein multiplied by point 56 and then you have this number you square it and you get a coefficient of variation of 8 3 4 in this case if we add an ingredient and we add an ingredient that it's not the best ingredient is probably one of the worst ingredient because it has the highest coefficient of variation it's still it has a tan coefficient of variation and we add a small amount of it so if you one would think but we are going to worsen the overall data no due to the fact that we are working on the square root we are improving variability we are reducing the variability so protein variability goes away we gets lower due to the addition of another ingredient of the worst quality from the point of view of variability and that in that diet pretty weird let's say at least not something that we would expect but that is explains why traditionally we have been making diets with many ingredients because many ingredients even off the results even without knowing we know that you use a little bit of this a little bit of them you will even out the result but so adding one more ingredient to your diet can decrease variability even if you are using an ingredient that is highly highly variable we saw this in the beginning of the year 2000 when we were deprived of the animal protein and we went it was very difficult to convince the feed Millers in Italy that they could they needed to substitute they the meat and bone with another ingredient and so we simply reduced a nutrient with the ingredients and we went to almost only soybean meal diet and the result was that we saw an increased variability in our diets all the nutrients increase their variability and our results also little bit suffered in the sense that the averages were good but they were opened up much more flatter distribution of results in technique in technical terms in the field the problem in using more ingredients or adding one ingredient to your list beside the practical side is that do you know your ingredient well enough because the problem is are you having the same level of certainty about this new ingredient how as you have for the other one you've been using because you cannot put things that you know perfectly together with things that you know nothing about this is a very complicated issue because you should know at least what you're dealing with not just put something in thinking that you know it but having very little ideas about it so yes it can work but your quality control Quality Assurance system must be very adequate and must be able to give you the data you need and are we sure that we are using our analytical capacities the money we spend in controlling in the right way how are we spending them in the that's just to check the fact that we are controlling the ingredients that are creating most of the variation that that's the key we must not spend money in things that are irrelevant to the variation this is a long-standing battle I once was running a filmy lab and my my lab technicians were very clever but they were very they filled themselves a scientist and so they were they were doing day-to-day controls but they were thinking in terms of doing analysis for micro elements so how much molybdenum was in the ingredients or those kinds of things and I had a very huge quarrel one day with them because I said what we need what I need from you his numbers that give me the chance to take a decision I don't need numbers for knowledge I need numbers for decision knowledge is good I can use historical data but when I use them once a year I've used them it's very difficult for me to put them to use every day while I need everyday numbers to tell me that if I have to accept or reject that load and that makes a very very big difference in terms of your variability management so are we concentrating the world of our scarce because they are always cares and I did to get control in the proper way because this is the very big question we have to make ourselves then practically do we have the place for a new ingredient have we got the bean well yes and no but depends let's say the advantage of one bin is so large that is you probably can find it and it's very depending on where you are in the world for example and the next example I will make is much more practical than one would think for example under our condition in Italy because normally we have very very big scales very long scales and normally people in the feed mills have at least two bins for any large ingredient in order to dose them at the extreme of the scale to prevent the scale from creating problems of poor way so two ways to practical ways to deal with that with the thing one is over formulation okay so you want ten you ask hundred and ten you can do it two ways one surely increasing the nutritional concentration of the diet and two discounting the ingredients the key nutrients of certain ingredients by a percentage of the standard deviation well the first one I'd rather forget because there is no way it no way it's it's apodictic you you cannot check it what why well you decide for 3% 5% 7% minus 10% whatever I mean there's no control on it it's just a good sense but it's very arbitrary and how can you say that we are not over formulating now because the problem is that if your reference is the standards of a breeding company well it depends where you are what is the quality of the chicks you get I mean it's so I'd rather think in terms of discounting but this counting is expensive of course because you decrease the value of an ingredient okay but certain key ingredients you do a decrease of their specific value in certain areas it's also interesting because this gives you the chance of in a way be always within the low limit slow ask use to be within a certain variation around what you declare on the label so in this way you're sure you're always you have a higher probability of being within the label value you declare normally what is applied is a minus 1/2 start and deviation out of the average you you have so that you can get to a point where you have about 67% I eat 2/3 of the samples that are better do what you expect of course one standard deviation would be better that would be a lot more more than 80% of the samples but that is going to be extremely expensive but let's say that the minus half a standard deviation is pretty practiced let's put it this way but then we take for granted that all ingredients are normally distributed and their normal distribution is pretty unskilled with the symmetric are our ingredients really so normal the problem is that I don't think so this is a two years of soybean meal this is Argentinian soybean meal from me the Italian feed compounder so it's 2 from January 1st 2017 to December 31st 2018 ok this is protein ok and you see their variability if we turn the same the other way around this is their distribution in terms of frequency at different protein levels okay and you see here indicated minus one minus two standard deviation please note that is not very symmetrical you have a much steeper side on the slide and drops much more here so the numerosity of samples and the minus one is much bigger than the numerosity and the plus one start the deviation if we check another soybean meal same same company this is a different this is hi-pro italian crush three different oil mills you see quite a an interesting variability about these things if we do the same three rebuilding of the numbers you see that here the structure is okay forty-eight average but here is completely the reverse you have many more data points above the average then below the evidence in the high side that of course here it's the reverse so what does this mean well this means that if you if you take a let's say a margin and you reduce by half a standard deviation your value the value you're using for soybean meal in general and you apply it to this soybean meal and this soybean meal so you get here or you get here well there's a substantial difference in the probability of over formulating how much excess protein you will be making by using soybean meal so in this case you will not be so oh you will be over formulating but the amount of excess protein would be not so high in this case this is going to be definitely higher and we also have problems in all the formulation we don't want so much excess nitrogen in our diets especially in certain diets so there's also the other side of the moon that we have to take into consideration and it's difficult to put the two things together so a more desirable approach for example would be to reduce ingredient variation by keeping the thing separate increasing the number of ingredients in a way but in the sense that we are considering them different either we separate the same ingredient on nutrient content we decide which is the key nutrient and then we we separate the ingredients we have a break-even point or where we decide above that its product a below that is product B or by simple physical separation we we randomly throw the same ingredient in a number of bins and then we pick the ingredient from in the same amount from all these bins what is the effect in practical life let's make this comparison of protein content in five different supplier of soybean meal so these are the five suppliers this is data from Duncan in very very old data but anyway it's not different if you take this numbers okay each each supplier has its own me in a zone standard deviation someone is better someone is worse okay if we we try and recombine the same okay keeping them separate and not separate meaning I I check the protein and I put a protein lower than the average in being a and protein higher than the average in decided in bin B and then I take 50% out of each in doing the diet or I use them randomly so with separation yes with separation and without separation you see that the standard deviation without separation is pretty substantial it's calculated according to to the equation I showed you before and also the variability for lies in for total sulphur in soybean meal and in the diet and if you separate in terms of plus and minus the average you see that you can drop your variability in the soybean meal by half and the same in the diet so it's definitely very interesting it's a very good approach it pays back and whenever you can do it you'd better do it it's a very easy thing especially today with nir technology if an ir would be used in the feed mill as a production tool and not in the lab as an analytical tool and that is a very big problem in some areas mine for example where the labs keep they and I are as a strictly as a analytical thing that it is not and there's a lot of companies that are not really using NIR for it's worth that mainly for me would be this kind of thing whoa what does it mean okay this is you have to test soybean meal and decide if it's quality air quality me but what happens if here what happens if instead of doing this we simply increase the number of Bin's okay so if we do separation without checking just simply putting the things one track in one bin one track in the other bin we have a definite improvement but it's only half of what we get if we do the analysis if you do the control if we have four random bins we get to more or less the same level of variability then we have by checking the data so it's a question of resources which kind of resources you want to have you can you spend the NIR analysis or the bins rebelde means no use for our analysis don't you have the bins use the analysis but you can do something the problem is that you have to recognize the need for you to do something and that is the biggest problem then there's other things and they here we go onto very very slimy area optimization tools yes there's optimization tools there are you can do almost everything they are not just a usual linear programming because they include in the variability data foreign group or ingredients so you need to have the variability data for each nutrient for each ingredient or at least for the ingredients you want to control and for the nutrient in the ingredients you want to control you need to decide what is the probability of success you want ie how many times out of a hundred you want to be within the numbers you you set as a requirement in your LP and you have to know your analytical error or data let's say that most of these things are totally unknown or very difficult to obtain and very difficult to manage and very difficult to maintain said this that explains why the software is not being a big success I can show you some idea of how it works here you see this is an optimization screen you see there's an area where it says standard deviations probability deviations analytical errors this is the area where you will add these numbers in order to be able to manage these things so first standard deviation you have to put in the standard deviation for in this case I used just protein and energy young shake energy okay this is a variation in absolute term so it's 15 kilocalories 15 kilo calories and point two okay so this is normal chord let's say and this is a highly variable chord for energy a higher variable chord for energy this is soybean meal we have the higher variability for protein okay so you have I have two ingredients one normal and one more variable for one of the nutrients this can happen difference of bias can have different variabilities then you ask what is the probability you want to see so i here said that i want to be sure at least 80% of the times that the result i get are within my limits for protein but they accept that only 50% is for energy fifty percent means that is not working it's doing nothing because normally when you're using averages you are 50% above a 50 percent below so asking 50 means if you put 50 everywhere you have no optimization difference against your standard linear programming optimization that will always give you data 50% about a 50% below then I set a specification and a very very simple specification where I put three two nutrients okay protein hmm from 18 to 20% energy from 3,000 to 3,250 lorries with the target probability of 80 and 50 as I showed before and this has been optimized at a certain cost that you see here in green okay so this diet here is this is the screen that shows you all the probability data the following on will show you the ingredients but anyway this diet brings with it a TA here this diet brings with it 72% corn 27.2 soybean meal and that's it okay with these volumes with the these nutrients that you see here with where the arrows are and this for a total cost that is this one okay and we've been using the good low variability product okay if we substitute the good variability product that you see here up that you see up here up here these are the good ones with the bad ones with a more variable ones you see that the thing changes okay you see that the first one I've been stopped at zeroed so they don't use it these are the different values so we are using more variable ingredients okay and the result is more expensive so the story is that whenever you do you're adding limits you're adding constraints and these conflicts cost money but the problem is that is extremely difficult for you to forecast what is going to be the money you're going to spend within this system yes ma'am okay we're almost there so you see tiny differences can give you interesting differences in monetary terms well the problem is that it's extremely difficult to forecast what you are going to see and I agree that we'd need a nonlinear approach to the problem but the problem of a nonlinear approach is that it's not linear so you cannot make prediction out of it it's a black box you put numbers in rattle and then you have numbers out and then you change something that is in the four decimal of an ingredient rattle and you have something that's completely different and you cannot make any projection of what we'll be having and that's hell for us so optimization tool are available on the software market but they need a for assessment of your data can you imagine some inconsistency that you can really skip in linear programming can create you have up here it's very complex to use and to put into use and it's absolutely impossible I said difficult but it's an understatement it's absolutely impossible to use in everyday life in everyday optimization absolutely impossible other one unless you have a huge amount of resources delivered into it because that that's easy okay no problem if you have unlimited resources but we are not in a place or in a world where there's than limited resources anyway it is an excellent what if tool but you have to have money time in data and data in time very often are the same thing as we all know data is of most valuable assets we have okay but it's not for free the more data you have the easier is to control variation in an effective way but data is costing you money resources and organization it's very very difficult and from my point of view variation in the nutrients in the feed is just one of the ring in the production chain and my background as an integration nutritionist risk is clearly apparent here so we have to be aware of the weakest ring in the chain or the lowest state in the barrel because that is what first we have to meet if we took take a look again at the feed flow and the thing we made here in green you see all things that you are the nutritionist problem okay also the rounding error in a way but all the rest is still having a hue impact and it's out of our control so I think that we need a much more systemic approach to variability in the production chain not just nutrient variability management it's just part of the problem and we of course need to tackle it but don't dare to think that you saw the problem because you will never be able to solve it and you will always have variation issues not coming from you the feed has this big big problem that is always the culprit because it's costing money but one thing that we should try as hard as we can is to convince our people that ok we are responsible for cost but we are not the cause of everything said that you're all invited to the next ESPN in Italy in 2021 in remaining thank you very much