Issue 17472: Attributes to be added to Measurement (smm-rtf) Source: Cordys (Mr. Henk de Man, hdman(at)cordys.com) Nature: Uncategorized Issue Severity: Summary: Add the following attributes to Measurement, to enable expression of measurement precision, and stochastic aspects of measurements: “confidence interval upper bound”, “confidence interval lower bound”, “confidence” (being a probability) and “distribution”(string). Resolution: Revised Text: Actions taken: July 13, 2012: received issue Discussion: End of Annotations:===== s is issue # 17472 From: Henk de Man To: Juergen Boldt , "issues@omg.org" , "smm-rtf@omg.org" Subject: RE: issues 17472 - 17475 -- SMM RTF issues Thread-Topic: issues 17472 - 17475 -- SMM RTF issues Thread-Index: AQHNYRB5ZHhwo3QPgUiSoRbKmbLbHZcpGa+g Date: Sat, 14 Jul 2012 18:40:01 +0000 Accept-Language: en-US X-MS-Has-Attach: X-MS-TNEF-Correlator: x-originating-ip: [10.24.11.7] X-OriginalArrivalTime: 14 Jul 2012 18:40:02.0870 (UTC) FILETIME=[1490ED60:01CD61F0] With respect to #17472: These attributes seem better fit for Measure, the method of measurement. A measurement is a single data point obtained by applying a measure. The measure may have precision, confidence bounds and a distribution. With respect to #17473: Each of these types seem to be categories of measures. Perhaps SMM should include some built-in measure categories such as Estimators, Simulators and Benchmarks. With respect to #17474: Measures can be equated. If there are multiple, equally good direct measures for a given characteristic then they are be directly associated by the measure equivalence relationship. With respect to #17475: I.m good with this. From: Juergen Boldt [mailto:juergen@omg.org] Sent: Friday, July 13, 2012 10:59 AM To: issues@omg.org; smm-rtf@omg.org Subject: issues 17472 - 17475 -- SMM RTF issues This is issue # 17472 From: Henk de Man To: Larry Hines Cc: Juergen Boldt , "issues@omg.org" , "smm-rtf@omg.org" , Henk de Man , Alain Picard , Pete Rivett , Arne Berre , "fred.a.cummins" X-Gm-Message-State: ALoCoQksqIh7GVwAF6Mewh9JrqjgeeMHdxauYBrS5eXbuTthMImmfhusrd7koZaMBsSx94Rdlj8F Larry, See below. On Sat, Jul 14, 2012 at 8:40 PM, Larry Hines wrote: With respect to #17472: These attributes seem better fit for Measure, the method of measurement. A measurement is a single data point obtained by applying a measure. The measure may have precision, confidence bounds and a distribution. [hdm] ÂWith respect to #17472: It is not correct to state that a measurement is a single data point. See e.g.ÂHubbard, Douglas W., How to Measure Anything, Finding the Value of âIntangiblesâ in Business, John Wiley & Sons, New Jersey, 2010. His main point is, that, in business measurements, just point measures aren't good measures. Often the point measure is not known or is not 100 % trustable. It is very important to be able to state measurement (so the result of measurement) as something that resides in a confidence interval, according to a certain level of confidence. ÂA measurement can be a point value. But it should also be possible to state a measurement as e.g. a confidence interval, with a confidence value. So-far about confidence interval and confidence. Now about distribution (stochastic distribution): A Monte Carlo measurement (i.e. outcome of Multi Carlo simulation or experiment) is typically expressed as a set of values that can be best described by a distribution curve. Stochastic enabling of measurement is important. But the distribution itself can not very well be defined as part of the measure, because the measure is re-usable in different contexts, and the distribution will often be specific to that context.   With respect to #17473: Each of these types seem to be categories of measures. Perhaps SMM should include some built-in measure categories such as Estimators, Simulators and Benchmarks. [hdm] ÂWith respect to #17473: It is not about the measure, but about the observation. It is about a specification the type of the observation. Note that the same measure can be used for for estimated measurement, as-is measurement, simulated measurement.  With respect to #17474: Measures can be equated. If there are multiple, equally good direct measures for a given characteristic then they are be directly associated by the measure equivalence relationship. [hdm]ÂWith respect to #17474.ÂThis is not about equivalence. ÂIt is like method overloading in Java ... When the same direct measure is applied in different contexts, slightly different or additional arguments (parameters) might be required in its operation. So, actually, more than one operation would be required. During discussions with Alain Picard this came out as the best solution.  With respect to #17475: Iâm good with this.   From: Juergen Boldt [mailto:juergen@omg.org] Sent: Friday, July 13, 2012 10:59 AM To: issues@omg.org; smm-rtf@omg.org Subject: issues 17472 - 17475 -- SMM RTF issues  This is issue # 17472ÂÂÂFrom: Henk de Man To: Henk de Man CC: Juergen Boldt , "issues@omg.org" , "smm-rtf@omg.org" , Alain Picard , Pete Rivett , Arne Berre , "fred.a.cummins" Subject: RE: issues 17472 -- SMM RTF issues Thread-Topic: issues 17472 -- SMM RTF issues Thread-Index: AQHNZCP79dR81qHE+0yXusLumPpeXA== Date: Tue, 17 Jul 2012 13:56:37 +0000 Accept-Language: en-US X-MS-Has-Attach: X-MS-TNEF-Correlator: x-originating-ip: [10.24.11.71] X-OriginalArrivalTime: 17 Jul 2012 13:56:38.0850 (UTC) FILETIME=[FC9EA220:01CD6423] Where does the value of .certain level of confidence. come from? It comes from a measure of confidence which is a distinct measure from the original measure. The certain level of confidence is a measurement which is distinct from the original measurement. They are indeed related measurements and it would be nice if SMM or another meta-model captured that relationship. We inserted the demonstration of trending data in SMM to answer a similar issue. Often the trend of the measurements is more important that the absolute value of the individual measurements. But that doesn.t change what a measurement is, namely a single data point. If I were to say that it.s 100 degrees F give or take 2 degrees F, I would be asserting two measurements, 100 degrees F and 2 degrees F, and I would be asserting a relationship between the two measurements. The same applies to a distribution of measurements. I agree that is should be possible to state a measurement with respect to a confidence interval. But that relationship is a relationship between measurements of different measures. Stating it as an attribute is a mistake. From: Henk de Man [mailto:hdman@cordys.com] Sent: Tuesday, July 17, 2012 8:01 AM To: Larry Hines Cc: Juergen Boldt; issues@omg.org; smm-rtf@omg.org; Henk de Man; Alain Picard; Pete Rivett; Arne Berre; fred.a.cummins Subject: Re: issues 17472 - 17475 -- SMM RTF issues Larry, See below. On Sat, Jul 14, 2012 at 8:40 PM, Larry Hines wrote: With respect to #17472: These attributes seem better fit for Measure, the method of measurement. A measurement is a single data point obtained by applying a measure. The measure may have precision, confidence bounds and a distribution. [hdm] With respect to #17472: It is not correct to state that a measurement is a single data point. See e.g. Hubbard, Douglas W., How to Measure Anything, Finding the Value of .Intangibles. in Business, John Wiley & Sons, New Jersey, 2010. His main point is, that, in business measurements, just point measures aren't good measures. Often the point measure is not known or is not 100 % trustable. It is very important to be able to state measurement (so the result of measurement) as something that resides in a confidence interval, according to a certain level of confidence. A measurement can be a point value. But it should also be possible to state a measurement as e.g. a confidence interval, with a confidence value. So-far about confidence interval and confidence. Now about distribution (stochastic distribution): A Monte Carlo measurement (i.e. outcome of Multi Carlo simulation or experiment) is typically expressed as a set of values that can be best described by a distribution curve. Stochastic enabling of measurement is important. But the distribution itself can not very well be defined as part of the measure, because the measure is re-usable in different contexts, and the distribution will often be specific to that context. With respect to #17473: Each of these types seem to be categories of measures. Perhaps SMM should include some built-in measure categories such as Estimators, Simulators and Benchmarks. [hdm] With respect to #17473: It is not about the measure, but about the observation. It is about a specification the type of the observation. Note that the same measure can be used for for estimated measurement, as-is measurement, simulated measurement. With respect to #17474: Measures can be equated. If there are multiple, equally good direct measures for a given characteristic then they are be directly associated by the measure equivalence relationship. [hdm] With respect to #17474. This is not about equivalence. It is like method overloading in Java ... When the same direct measure is applied in different contexts, slightly different or additional arguments (parameters) might be required in its operation. So, actually, more than one operation would be required. During discussions with Alain Picard this came out as the best solution. With respect to #17475: I.m good with this. From: Juergen Boldt [mailto:juergen@omg.org] Sent: Friday, July 13, 2012 10:59 AM To: issues@omg.org; smm-rtf@omg.org Subject: issues 17472 - 17475 -- SMM RTF issues This is issue # 17472 From: Henk de Man To: Larry Hines Cc: Juergen Boldt , "issues@omg.org" , "smm-rtf@omg.org" , Alain Picard , Pete Rivett , Arne Berre , "fred.a.cummins" , Henk de Man X-Gm-Message-State: ALoCoQnh/9kytF20dfxhexMncos3DxjVHqxdsE7/jt1GT4vdVZweogl0E4D45c+770386doYnOs0 Larry, See below. On Tue, Jul 17, 2012 at 3:56 PM, Larry Hines wrote: Where does the value of âcertain level of confidenceâ come from? [hdm] It is as good as the business analyst can make it. Often he/she doesn't have a point value, but just knows that e.g. the value is between "this" and "that" with x % probability ("confidence"). In other words: there's a degree of uncertainty left, and the only thing the measure does is narrowing down the uncertainty to the bounds of the confidence interval. But note again that these bounds aren't coming by measure (library), but are outcome of the mind of the business analyst on the spot. See the many examples inÂHubbard, Douglas W.,ÂHow to Measure Anything, Finding the Value of âIntangiblesâ in Business, John Wiley & Sons, New Jersey, 2010.  It comes from a measure of confidence which is a distinct measure from the original measure.  [hdm] No, no. See above. ÂPrecision of a measurement is not given by the measure.. It is inherently or intrinsically part of the measurement itself. This is important. The certain level of confidence is a measurement which is distinct from the original measurement. [hdm] No, because in many of these situations there is no "original measurement". There are just these bounds. And these form the measurement. Don't make each bound a separate measurement, because then we get burdened with so Âmany objects in the model. And note that for all of these things the measure is just the measure: its name, its characteristic that is measures, its unit, etc. Care for maintainability and purity and re-usability of the library. They are indeed related measurements and it would be nice if SMM or another meta-model captured that relationship.  We inserted the demonstration of trending data in SMM to answer a similar issue. Often the trend of the measurements is more important that the absolute value of the individual measurements. But that doesnât change what a measurement is, namely a single data point. [hdm] The proposal is to add e.g. confidence interval bounds and probability to the point value, to have the measurement being expressed by a couple of things, rather than just that point value. And note again: often a point value cannot be given, but confidence interval can be. See again Hubbard, who is very major on this.   If I were to say that itâs 100 degrees F give or take 2 degrees F, I would be asserting two measurements, 100 degrees F and 2 degrees F, and I would be asserting a relationship between the two measurements. The same applies to a distribution of measurements. [hdm] Note that what we primarily want to achieve with the distribution is: an expression of outcome of measurement. For instance, in a MonteCarlo simulation, there are 10.000 (e.g.) measurements per each "leave" measure (the direct ones). All of these are aggregated up (so to say), taking into account the so many binary, rescale and collective effects of the other measures, and then "at the top", all these 10.000 measurements come out (per each "top" measure). Consider a VDML model with, say 100 measured characterstics. A simulation engine would generate (in its run-time database or memory) 100 * 10.000 = 1.000.000 measurements.. What we want to import back in the VDML model (the model of the business system or business system scenario), is not 1.000.000 measurements. Of course not. We want an observation for the entire simulation, containing then just 100 measurements, whereby each measurement has a distribution (based on curve fitting, for instance), that describes the outcome. That's also what Monte Carlo does: rendering distributions of simulated values of such characteristics. So, the distributions, as part of measurement, is not meant as direct measure operation, but primarily as measurement outcome. Describing a distribution of values, rather than just one point value. And what you say, storing the all the so-many from-simulation-memory values as separate measurement objects, would not be feasible, as it would load a VDML model with 1.000.000 objects... Now, given that we need a refinement of a measurement with distribution, we can further argue about what this means at the "bottom" (the direct measurements, the "leaves"): do we want an operation to define that the simulation would draw from a distribution, so, do we want to implement a distribution as an operation ? I would say: Given that we want a distribution, as outcome, as part of a measurement, why can't we re-use this same distribution to "teach" the simulation engine on the "leaves"? If you would insist on an operation for that, ok, but then only when you allow for 0..* operations on a measure, otherwise we get that duplication-of-library-nightmare..   I agree that is should be possible to state a measurement with respect to a confidence interval. But that relationship is a relationship between measurements of different measures. Stating it as an attribute is a mistake. [hdm] We need to be practical. We need to care for maintainability and scalability. See above.    From: Henk de Man [mailto:hdman@cordys.com] Sent: Tuesday, July 17, 2012 8:01 AM To: Larry Hines Cc: Juergen Boldt; issues@omg.org; smm-rtf@omg.org; Henk de Man; Alain Picard; Pete Rivett; Arne Berre; fred.a.cummins Subject: Re: issues 17472 - 17475 -- SMM RTF issues  Larry,  See below. On Sat, Jul 14, 2012 at 8:40 PM, Larry Hines wrote: With respect to #17472: These attributes seem better fit for Measure, the method of measurement. A measurement is a single data point obtained by applying a measure. The measure may have precision, confidence bounds and a distribution. [hdm] ÂWith respect to #17472: It is not correct to state that a measurement is a single data point. See e.g.ÂHubbard, Douglas W., How to Measure Anything, Finding the Value of âIntangiblesâ in Business, John Wiley & Sons, New Jersey, 2010. His main point is, that, in business measurements, just point measures aren't good measures. Often the point measure is not known or is not 100 % trustable. It is very important to be able to state measurement (so the result of measurement) as something that resides in a confidence interval, according to a certain level of confidence. ÂA measurement can be a point value. But it should also be possible to state a measurement as e.g. a confidence interval, with a confidence value. So-far about confidence interval and confidence. Now about distribution (stochastic distribution): A Monte Carlo measurement (i.e. outcome of Multi Carlo simulation or experiment) is typically expressed as a set of values that can be best described by a distribution curve. Stochastic enabling of measurement is important. But the distribution itself can not very well be defined as part of the measure, because the measure is re-usable in different contexts, and the distribution will often be specific to that context.   With respect to #17473: Each of these types seem to be categories of measures. Perhaps SMM should include some built-in measure categories such as Estimators, Simulators and Benchmarks. [hdm] ÂWith respect to #17473: It is not about the measure, but about the observation. It is about a specification the type of the observation. Note that the same measure can be used for for estimated measurement, as-is measurement, simulated measurement.  With respect to #17474: Measures can be equated. If there are multiple, equally good direct measures for a given characteristic then they are be directly associated by the measure equivalence relationship. [hdm]ÂWith respect to #17474.ÂThis is not about equivalence. ÂIt is like method overloading in Java ... When the same direct measure is applied in different contexts, slightly different or additional arguments (parameters) might be required in its operation. So, actually, more than one operation would be required. During discussions with Alain Picard this came out as the best solution.  With respect to #17475: Iâm good with this.   From: Juergen Boldt [mailto:juergen@omg.org] Sent: Friday, July 13, 2012 10:59 AM To: issues@omg.org; smm-rtf@omg.org Subject: issues 17472 - 17475 -- SMM RTF issues  This is issue # 17472ÂÂÂFrom: Henk de Man To: Henk de Man CC: Juergen Boldt , "issues@omg.org" , "smm-rtf@omg.org" , Alain Picard , Pete Rivett , Arne Berre , "fred.a.cummins" Subject: RE: issues 17472 -- SMM RTF issues Thread-Topic: issues 17472 -- SMM RTF issues Thread-Index: AQHNZCP79dR81qHE+0yXusLumPpeXJcvfDUA///A8SA= Date: Wed, 18 Jul 2012 19:30:24 +0000 Accept-Language: en-US X-MS-Has-Attach: X-MS-TNEF-Correlator: x-originating-ip: [10.64.26.13] X-OriginalArrivalTime: 18 Jul 2012 19:30:25.0215 (UTC) FILETIME=[C7AD2CF0:01CD651B] Henk, A measure is an evaluation process that assigns a comparable numeric or symbolic value to an entity in order to characterize a selected quantity or trait of the entity. The evaluation process may be a mathematically formula applied to given values, the application of a tool to an entity or the fuzzy logic of a business analyst. It can be any process which evaluates an entity to produce a comparable value. Such evaluations are a large part of what business analyst are paid for. In fact, if the business analyst process is not defined (at least by a name) as a measure in SMM then his or her measurements cannot be represented in SMM. Width is inherently part of a line. It.s value is obtained by measuring the line. Precision and accuracy are likewise obtained by evaluation processes. Precision: Precision is the degree to which the correctness of a measurement is expressed. Often the precision of a measure is identical to its unit of measure. Examining the ruler that I keep in my desk I see its precision is one tenth of an inch on one side and a millimeter of the other. One tenth of an inch is a measurement inherent to measuring with that side of my ruler. Since precision is not always identical unit of measure, adding precision as an optional attribute of dimensional measure. Accuracy: Based up my experience I would evaluate the accuracy of Google Maps as very good. .Very good. is better than .good. but worse than .perfect.. If one wants a numeric value for the accuracy of Google Maps, we could select a number of places for which we have known longitudes and latitudes and we test Google Maps. The average of the absolute divergence from the known location would be one measure of Google Map.s accuracy. A range over a linear space is given by two values. One can compare the minimums, maximums, midpoints and lengths of ranges. But the ranges themselves are not directly comparable. That is, given ranges R1 and R2, R1 < R2 is not determined. Consequently, ranges are not measurements. One can, however, find a lower bound and a upper bound. That.s two measurements. The two measurements have the same measurand and are under the same observation. Yes, there is some inconvenience here. But what you want to represent can be represented correctly in SMM. With respect to scalability, if all one wants is the top level 100 measurements without any of the myriad of measurement aggregated to determine them, then that is all that SMM requires. The lower level measurements are not required. One can cut it off at any level. If one wants the top 100 distributions then one only needs the 100 sets of measurements that imply those 100 distributions, not the million measurements that went into their calculation. My basic point is that we should not lose the definitions of measure and measurement. If we lose those definitions then we have nothing. I.m tempted to propose an interval measure but that.s a can of worms to be thought out carefully. From: Henk de Man [mailto:hdman@cordys.com] Sent: Wednesday, July 18, 2012 11:17 AM To: Larry Hines Cc: Juergen Boldt; issues@omg.org; smm-rtf@omg.org; Alain Picard; Pete Rivett; Arne Berre; fred.a.cummins; Henk de Man Subject: Re: issues 17472 -- SMM RTF issues Larry, See below. On Tue, Jul 17, 2012 at 3:56 PM, Larry Hines wrote: Where does the value of .certain level of confidence. come from? [hdm] It is as good as the business analyst can make it. Often he/she doesn't have a point value, but just knows that e.g. the value is between "this" and "that" with x % probability ("confidence"). In other words: there's a degree of uncertainty left, and the only thing the measure does is narrowing down the uncertainty to the bounds of the confidence interval. But note again that these bounds aren't coming by measure (library), but are outcome of the mind of the business analyst on the spot. See the many examples in Hubbard, Douglas W., How to Measure Anything, Finding the Value of .Intangibles. in Business, John Wiley & Sons, New Jersey, 2010. It comes from a measure of confidence which is a distinct measure from the original measure. [hdm] No, no. See above. Precision of a measurement is not given by the measure.. It is inherently or intrinsically part of the measurement itself. This is important. The certain level of confidence is a measurement which is distinct from the original measurement. [hdm] No, because in many of these situations there is no "original measurement". There are just these bounds. And these form the measurement. Don't make each bound a separate measurement, because then we get burdened with so many objects in the model. And note that for all of these things the measure is just the measure: its name, its characteristic that is measures, its unit, etc. Care for maintainability and purity and re-usability of the library. They are indeed related measurements and it would be nice if SMM or another meta-model captured that relationship. We inserted the demonstration of trending data in SMM to answer a similar issue. Often the trend of the measurements is more important that the absolute value of the individual measurements. But that doesn.t change what a measurement is, namely a single data point. [hdm] The proposal is to add e.g. confidence interval bounds and probability to the point value, to have the measurement being expressed by a couple of things, rather than just that point value. And note again: often a point value cannot be given, but confidence interval can be. See again Hubbard, who is very major on this. If I were to say that it.s 100 degrees F give or take 2 degrees F, I would be asserting two measurements, 100 degrees F and 2 degrees F, and I would be asserting a relationship between the two measurements. The same applies to a distribution of measurements. [hdm] Note that what we primarily want to achieve with the distribution is: an expression of outcome of measurement. For instance, in a MonteCarlo simulation, there are 10.000 (e.g.) measurements per each "leave" measure (the direct ones). All of these are aggregated up (so to say), taking into account the so many binary, rescale and collective effects of the other measures, and then "at the top", all these 10.000 measurements come out (per each "top" measure). Consider a VDML model with, say 100 measured characterstics. A simulation engine would generate (in its run-time database or memory) 100 * 10.000 = 1.000.000 measurements.. What we want to import back in the VDML model (the model of the business system or business system scenario), is not 1.000.000 measurements. Of course not. We want an observation for the entire simulation, containing then just 100 measurements, whereby each measurement has a distribution (based on curve fitting, for instance), that describes the outcome. That's also what Monte Carlo does: rendering distributions of simulated values of such characteristics. So, the distributions, as part of measurement, is not meant as direct measure operation, but primarily as measurement outcome. Describing a distribution of values, rather than just one point value. And what you say, storing the all the so-many from-simulation-memory values as separate measurement objects, would not be feasible, as it would load a VDML model with 1.000.000 objects... Now, given that we need a refinement of a measurement with distribution, we can further argue about what this means at the "bottom" (the direct measurements, the "leaves"): do we want an operation to define that the simulation would draw from a distribution, so, do we want to implement a distribution as an operation ? I would say: Given that we want a distribution, as outcome, as part of a measurement, why can't we re-use this same distribution to "teach" the simulation engine on the "leaves"? If you would insist on an operation for that, ok, but then only when you allow for 0..* operations on a measure, otherwise we get that duplication-of-library-nightmare.. I agree that is should be possible to state a measurement with respect to a confidence interval. But that relationship is a relationship between measurements of different measures. Stating it as an attribute is a mistake. [hdm] We need to be practical. We need to care for maintainability and scalability. See above. From: Henk de Man [mailto:hdman@cordys.com] Sent: Tuesday, July 17, 2012 8:01 AM To: Larry Hines Cc: Juergen Boldt; issues@omg.org; smm-rtf@omg.org; Henk de Man; Alain Picard; Pete Rivett; Arne Berre; fred.a.cummins Subject: Re: issues 17472 - 17475 -- SMM RTF issues Larry, See below. On Sat, Jul 14, 2012 at 8:40 PM, Larry Hines wrote: With respect to #17472: These attributes seem better fit for Measure, the method of measurement. A measurement is a single data point obtained by applying a measure. The measure may have precision, confidence bounds and a distribution. [hdm] With respect to #17472: It is not correct to state that a measurement is a single data point. See e.g. Hubbard, Douglas W., How to Measure Anything, Finding the Value of .Intangibles. in Business, John Wiley & Sons, New Jersey, 2010. His main point is, that, in business measurements, just point measures aren't good measures. Often the point measure is not known or is not 100 % trustable. It is very important to be able to state measurement (so the result of measurement) as something that resides in a confidence interval, according to a certain level of confidence. A measurement can be a point value. But it should also be possible to state a measurement as e.g. a confidence interval, with a confidence value. So-far about confidence interval and confidence. Now about distribution (stochastic distribution): A Monte Carlo measurement (i.e. outcome of Multi Carlo simulation or experiment) is typically expressed as a set of values that can be best described by a distribution curve. Stochastic enabling of measurement is important. But the distribution itself can not very well be defined as part of the measure, because the measure is re-usable in different contexts, and the distribution will often be specific to that context. With respect to #17473: Each of these types seem to be categories of measures. Perhaps SMM should include some built-in measure categories such as Estimators, Simulators and Benchmarks. [hdm] With respect to #17473: It is not about the measure, but about the observation. It is about a specification the type of the observation. Note that the same measure can be used for for estimated measurement, as-is measurement, simulated measurement. With respect to #17474: Measures can be equated. If there are multiple, equally good direct measures for a given characteristic then they are be directly associated by the measure equivalence relationship. [hdm] With respect to #17474. This is not about equivalence. It is like method overloading in Java ... When the same direct measure is applied in different contexts, slightly different or additional arguments (parameters) might be required in its operation. So, actually, more than one operation would be required. During discussions with Alain Picard this came out as the best solution. With respect to #17475: I.m good with this. From: Juergen Boldt [mailto:juergen@omg.org] Sent: Friday, July 13, 2012 10:59 AM To: issues@omg.org; smm-rtf@omg.org Subject: issues 17472 - 17475 -- SMM RTF issues This is issue # 17472 From: Henk de Man To: Larry Hines Cc: Juergen Boldt , "issues@omg.org" , "smm-rtf@omg.org" , Alain Picard , Pete Rivett , Arne Berre , "fred.a.cummins" , Henk de Man X-Gm-Message-State: ALoCoQle133SRWpLniJaBP4bAMqnaHo5KN1FIDMMyCaNVttRvG1ywThWih4BzjamYxEah1yBJ+ez Larry, See attachment for detailed discussion on your topic below. I think I came into your direction, but this would require a couple of things to be clarified, see attachment. Can you react to it and see what you think ? Regards, Henk de Man On Wed, Jul 18, 2012 at 9:30 PM, Larry Hines wrote: Henk,  A measure is an evaluation process that assigns a comparable numeric or symbolic value to an entity in order to characterize a selected quantity or trait of the entity.  The evaluation process may be a mathematically formula applied to given values, the application of a tool to an entity or the fuzzy logic of a business analyst. It can be any process which evaluates an entity to produce a comparable value. Such evaluations are a large part of what business analyst are paid for. In fact, if the business analyst process is not defined (at least by a name) as a measure in SMM then his or her measurements cannot be represented in SMM.  Width is inherently part of a line. Itâs value is obtained by measuring the line. Precision and accuracy are likewise obtained by evaluation processes.  Precision: Precision is the degree to which the correctness of a measurement is expressed. Often the precision of a measure is identical to its unit of measure. Examining the ruler that I keep in my desk I see its precision is one tenth of an inch on one side and a millimeter of the other. One tenth of an inch is a measurement inherent to measuring with that side of my ruler. Since precision is not always identical unit of measure, adding precision as an optional attribute of dimensional measure.  Accuracy: Based up my experience I would evaluate the accuracy of Google Maps as very good. âVery goodâ is better than âgoodâ but worse than âperfectâ.  If one wants a numeric value for the accuracy of Google Maps, we could select a number of places for which we have known longitudes and latitudes and we test Google Maps. The average of the absolute divergence from the known location would be one measure of Google Mapâs accuracy.  A range over a linear space is given by two values. One can compare the minimums, maximums, midpoints and lengths of ranges. But the ranges themselves are not directly comparable. That is, given ranges R1 and R2, R1 < R2 is not determined. Consequently, ranges are not measurements.  One can, however, find a lower bound and a upper bound. Thatâs two measurements. The two measurements have the same measurand and are under the same observation. Yes, there is some inconvenience here. But what you want to represent can be represented correctly in SMM. With respect to scalability, if all one wants is the top level 100 measurements without any of the myriad of measurement aggregated to determine them, then that is all that SMM requires. ÂÂÂThe lower level measurements are not required. One can cut it off at any level.  If one wants the top 100 distributions then one only needs the 100 sets of measurements that imply those 100 distributions, not the million measurements that went into their calculation.  My basic point is that we should not lose the definitions of measure and measurement. If we lose those definitions then we have nothing.  Iâm tempted to propose an interval measure but thatâs a can of worms to be thought out carefully.  From: Henk de Man [mailto:hdman@cordys.com] Sent: Wednesday, July 18, 2012 11:17 AM To: Larry Hines Cc: Juergen Boldt; issues@omg.org; smm-rtf@omg.org; Alain Picard; Pete Rivett; Arne Berre; fred.a.cummins; Henk de Man Subject: Re: issues 17472 -- SMM RTF issues  Larry,  See below. On Tue, Jul 17, 2012 at 3:56 PM, Larry Hines wrote: Where does the value of âcertain level of confidenceâ come from? [hdm] It is as good as the business analyst can make it. Often he/she doesn't have a point value, but just knows that e.g. the value is between "this" and "that" with x % probability ("confidence"). In other words: there's a degree of uncertainty left, and the only thing the measure does is narrowing down the uncertainty to the bounds of the confidence interval. But note again that these bounds aren't coming by measure (library), but are outcome of the mind of the business analyst on the spot. See the many examples inÂHubbard, Douglas W.,ÂHow to Measure Anything, Finding the Value of âIntangiblesâ in Business, John Wiley & Sons, New Jersey, 2010.  It comes from a measure of confidence which is a distinct measure from the original measure.  [hdm] No, no. See above. ÂPrecision of a measurement is not given by the measure.. It is inherently or intrinsically part of the measurement itself. This is important. The certain level of confidence is a measurement which is distinct from the original measurement. [hdm] No, because in many of these situations there is no "original measurement". There are just these bounds. And these form the measurement. Don't make each bound a separate measurement, because then we get burdened with so Âmany objects in the model. And note that for all of these things the measure is just the measure: its name, its characteristic that is measures, its unit, etc. Care for maintainability and purity and re-usability of the library.   They are indeed related measurements and it would be nice if SMM or another meta-model captured that relationship.  We inserted the demonstration of trending data in SMM to answer a similar issue. Often the trend of the measurements is more important that the absolute value of the individual measurements. But that doesnât change what a measurement is, namely a single data point. [hdm] The proposal is to add e.g. confidence interval bounds and probability to the point value, to have the measurement being expressed by a couple of things, rather than just that point value. And note again: often a point value cannot be given, but confidence interval can be. See again Hubbard, who is very major on this.   If I were to say that itâs 100 degrees F give or take 2 degrees F, I would be asserting two measurements, 100 degrees F and 2 degrees F, and I would be asserting a relationship between the two measurements. The same applies to a distribution of measurements. [hdm] Note that what we primarily want to achieve with the distribution is: an expression of outcome of measurement. For instance, in a MonteCarlo simulation, there are 10.000 (e.g.) measurements per each "leave" measure (the direct ones). All of these are aggregated up (so to say), taking into account the so many binary, rescale and collective effects of the other measures, and then "at the top", all these 10.000 measurements come out (per each "top" measure). Consider a VDML model with, say 100 measured characterstics. A simulation engine would generate (in its run-time database or memory) 100 * 10.000 = 1.000.000 measurements.. What we want to import back in the VDML model (the model of the business system or business system scenario), is not 1.000.000 measurements. Of course not. We want an observation for the entire simulation, containing then just 100 measurements, whereby each measurement has a distribution (based on curve fitting, for instance), that describes the outcome. That's also what Monte Carlo does: rendering distributions of simulated values of such characteristics. So, the distributions, as part of measurement, is not meant as direct measure operation, but primarily as measurement outcome. Describing a distribution of values, rather than just one point value. And what you say, storing the all the so-many from-simulation-memory values as separate measurement objects, would not be feasible, as it would load a VDML model with 1.000.000 objects...  Now, given that we need a refinement of a measurement with distribution, we can further argue about what this means at the "bottom" (the direct measurements, the "leaves"): do we want an operation to define that the simulation would draw from a distribution, so, do we want to implement a distribution as an operation ? I would say: Given that we want a distribution, as outcome, as part of a measurement, why can't we re-use this same distribution to "teach" the simulation engine on the "leaves"? If you would insist on an operation for that, ok, but then only when you allow for 0..* operations on a measure, otherwise we get that duplication-of-library-nightmare..   I agree that is should be possible to state a measurement with respect to a confidence interval. But that relationship is a relationship between measurements of different measures. Stating it as an attribute is a mistake. [hdm] We need to be practical. We need to care for maintainability and scalability. See above.    From: Henk de Man [mailto:hdman@cordys.com] Sent: Tuesday, July 17, 2012 8:01 AM To: Larry Hines Cc: Juergen Boldt; issues@omg.org; smm-rtf@omg.org; Henk de Man; Alain Picard; Pete Rivett; Arne Berre; fred.a.cummins Subject: Re: issues 17472 - 17475 -- SMM RTF issues  Larry,  See below. On Sat, Jul 14, 2012 at 8:40 PM, Larry Hines wrote: With respect to #17472: These attributes seem better fit for Measure, the method of measurement. A measurement is a single data point obtained by applying a measure. The measure may have precision, confidence bounds and a distribution. [hdm] ÂWith respect to #17472: It is not correct to state that a measurement is a single data point. See e.g.ÂHubbard, Douglas W., How to Measure Anything, Finding the Value of âIntangiblesâ in Business, John Wiley & Sons, New Jersey, 2010. His main point is, that, in business measurements, just point measures aren't good measures. Often the point measure is not known or is not 100 % trustable. It is very important to be able to state measurement (so the result of measurement) as something that resides in a confidence interval, according to a certain level of confidence. ÂA measurement can be a point value. But it should also be possible to state a measurement as e.g. a confidence interval, with a confidence value. So-far about confidence interval and confidence. Now about distribution (stochastic distribution): A Monte Carlo measurement (i.e. outcome of Multi Carlo simulation or experiment) is typically expressed as a set of values that can be best described by a distribution curve. Stochastic enabling of measurement is important. But the distribution itself can not very well be defined as part of the measure, because the measure is re-usable in different contexts, and the distribution will often be specific to that context.   With respect to #17473: Each of these types seem to be categories of measures. Perhaps SMM should include some built-in measure categories such as Estimators, Simulators and Benchmarks. [hdm] ÂWith respect to #17473: It is not about the measure, but about the observation. It is about a specification the type of the observation. Note that the same measure can be used for for estimated measurement, as-is measurement, simulated measurement.  With respect to #17474: Measures can be equated. If there are multiple, equally good direct measures for a given characteristic then they are be directly associated by the measure equivalence relationship. [hdm]ÂWith respect to #17474.ÂThis is not about equivalence. ÂIt is like method overloading in Java ... When the same direct measure is applied in different contexts, slightly different or additional arguments (parameters) might be required in its operation. So, actually, more than one operation would be required. During discussions with Alain Picard this came out as the best solution.  With respect to #17475: Iâm good with this.   From: Juergen Boldt [mailto:juergen@omg.org] Sent: Friday, July 13, 2012 10:59 AM To: issues@omg.org; smm-rtf@omg.org Subject: issues 17472 - 17475 -- SMM RTF issues  This is issue # 17472ÂÂÂFrom: Henk de Man X-Google-DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=google.com; s=20120113; h=mime-version:in-reply-to:references:date:message-id:subject:from:to :cc:content-type:x-gm-message-state; bh=pEFIza2JSEhIJIyfaCb+vsObEXtNQ725oMkQF9taqyc=; b=kqrOkYmu6MrvrrNLxjJnKHsVjB1rn1aKxn+vAPtDPAmvy1K9r1a3e7YFt/KTzCCyCC aO3IeU4Fn4wLQpB1f2sXG7AHrhflCjnusjfu/h1di1yp9rGsbcRmCVfVYwqqc0AIs/dB HliCI/5PI6KueGAXBWXfr4ajeGs3LFI5C81BN/VRPS5xRK8KrwYGWQgq+3i5Y3D7x3Nn rUHn6sZ1O6agrwaEKScYjHyeKRG9LfaIYqk2UHoFWyY/1VWquw+YIhGb2CPxLUWX4LfQ ambUejAISrDCuOQHlARKqbsOlzHE4BHwIUR2rN0C06rureczn0xz+alOWmR+xB3ie3x+ k3ag== Date: Fri, 20 Jul 2012 18:24:11 +0200 Subject: Re: issues 17472 -- SMM RTF issues From: Henk de Man To: Larry Hines Cc: Juergen Boldt , "issues@omg.org" , "smm-rtf@omg.org" , Alain Picard , Pete Rivett , Arne Berre , "fred.a.cummins" , Henk de Man X-Gm-Message-State: ALoCoQlYkzfVXWXO6tdbA0N8glaj5NGN4yrGhKOG9GeC9skkCwiwMkzkKhcxGREYk9rBAsBRwq1x Larry, Are you good with the feedback in the attachment below? If you are, how can we make it feasible, given its open issue in it. For convenience, attached the document again. Regards, Henk de Man On Thu, Jul 19, 2012 at 6:30 PM, Henk de Man wrote: Larry, See attachment for detailed discussion on your topic below. I think I came into your direction, but this would require a couple of things to be clarified, see attachment. Can you react to it and see what you think ? Regards, Henk de Man On Wed, Jul 18, 2012 at 9:30 PM, Larry Hines wrote: Henk,  A measure is an evaluation process that assigns a comparable numeric or symbolic value to an entity in order to characterize a selected quantity or trait of the entity.  The evaluation process may be a mathematically formula applied to given values, the application of a tool to an entity or the fuzzy logic of a business analyst. It can be any process which evaluates an entity to produce a comparable value. Such evaluations are a large part of what business analyst are paid for. In fact, if the business analyst process is not defined (at least by a name) as a measure in SMM then his or her measurements cannot be represented in SMM.  Width is inherently part of a line. Itâs value is obtained by measuring the line. Precision and accuracy are likewise obtained by evaluation processes.  Precision: Precision is the degree to which the correctness of a measurement is expressed. Often the precision of a measure is identical to its unit of measure. Examining the ruler that I keep in my desk I see its precision is one tenth of an inch on one side and a millimeter of the other. One tenth of an inch is a measurement inherent to measuring with that side of my ruler. Since precision is not always identical unit of measure, adding precision as an optional attribute of dimensional measure.  Accuracy: Based up my experience I would evaluate the accuracy of Google Maps as very good. âVery goodâ is better than âgoodâ but worse than âperfectâ.  If one wants a numeric value for the accuracy of Google Maps, we could select a number of places for which we have known longitudes and latitudes and we test Google Maps. The average of the absolute divergence from the known location would be one measure of Google Mapâs accuracy.  A range over a linear space is given by two values. One can compare the minimums, maximums, midpoints and lengths of ranges. But the ranges themselves are not directly comparable. That is, given ranges R1 and R2, R1 < R2 is not determined. Consequently, ranges are not measurements.  One can, however, find a lower bound and a upper bound. Thatâs two measurements. The two measurements have the same measurand and are under the same observation. Yes, there is some inconvenience here. But what you want to represent can be represented correctly in SMM. With respect to scalability, if all one wants is the top level 100 measurements without any of the myriad of measurement aggregated to determine them, then that is all that SMM requires. ÂÂÂThe lower level measurements are not required. One can cut it off at any level.  If one wants the top 100 distributions then one only needs the 100 sets of measurements that imply those 100 distributions, not the million measurements that went into their calculation.  My basic point is that we should not lose the definitions of measure and measurement. If we lose those definitions then we have nothing.  Iâm tempted to propose an interval measure but thatâs a can of worms to be thought out carefully.  From: Henk de Man [mailto:hdman@cordys.com] Sent: Wednesday, July 18, 2012 11:17 AM To: Larry Hines Cc: Juergen Boldt; issues@omg.org; smm-rtf@omg.org; Alain Picard; Pete Rivett; Arne Berre; fred.a.cummins; Henk de Man Subject: Re: issues 17472 -- SMM RTF issues  Larry,  See below. On Tue, Jul 17, 2012 at 3:56 PM, Larry Hines wrote: Where does the value of âcertain level of confidenceâ come from? [hdm] It is as good as the business analyst can make it. Often he/she doesn't have a point value, but just knows that e.g. the value is between "this" and "that" with x % probability ("confidence"). In other words: there's a degree of uncertainty left, and the only thing the measure does is narrowing down the uncertainty to the bounds of the confidence interval. But note again that these bounds aren't coming by measure (library), but are outcome of the mind of the business analyst on the spot. See the many examples inÂHubbard, Douglas W.,ÂHow to Measure Anything, Finding the Value of âIntangiblesâ in Business, John Wiley & Sons, New Jersey, 2010.  It comes from a measure of confidence which is a distinct measure from the original measure.  [hdm] No, no. See above. ÂPrecision of a measurement is not given by the measure.. It is inherently or intrinsically part of the measurement itself. This is important. The certain level of confidence is a measurement which is distinct from the original measurement. [hdm] No, because in many of these situations there is no "original measurement". There are just these bounds. And these form the measurement. Don't make each bound a separate measurement, because then we get burdened with so Âmany objects in the model. And note that for all of these things the measure is just the measure: its name, its characteristic that is measures, its unit, etc. Care for maintainability and purity and re-usability of the library.   They are indeed related measurements and it would be nice if SMM or another meta-model captured that relationship.  We inserted the demonstration of trending data in SMM to answer a similar issue. Often the trend of the measurements is more important that the absolute value of the individual measurements. But that doesnât change what a measurement is, namely a single data point. [hdm] The proposal is to add e.g. confidence interval bounds and probability to the point value, to have the measurement being expressed by a couple of things, rather than just that point value. And note again: often a point value cannot be given, but confidence interval can be. See again Hubbard, who is very major on this.   If I were to say that itâs 100 degrees F give or take 2 degrees F, I would be asserting two measurements, 100 degrees F and 2 degrees F, and I would be asserting a relationship between the two measurements. The same applies to a distribution of measurements. [hdm] Note that what we primarily want to achieve with the distribution is: an expression of outcome of measurement. For instance, in a MonteCarlo simulation, there are 10.000 (e.g.) measurements per each "leave" measure (the direct ones). All of these are aggregated up (so to say), taking into account the so many binary, rescale and collective effects of the other measures, and then "at the top", all these 10.000 measurements come out (per each "top" measure). Consider a VDML model with, say 100 measured characterstics. A simulation engine would generate (in its run-time database or memory) 100 * 10.000 = 1.000.000 measurements.. What we want to import back in the VDML model (the model of the business system or business system scenario), is not 1.000.000 measurements. Of course not. We want an observation for the entire simulation, containing then just 100 measurements, whereby each measurement has a distribution (based on curve fitting, for instance), that describes the outcome. That's also what Monte Carlo does: rendering distributions of simulated values of such characteristics. So, the distributions, as part of measurement, is not meant as direct measure operation, but primarily as measurement outcome. Describing a distribution of values, rather than just one point value. And what you say, storing the all the so-many from-simulation-memory values as separate measurement objects, would not be feasible, as it would load a VDML model with 1.000.000 objects...  Now, given that we need a refinement of a measurement with distribution, we can further argue about what this means at the "bottom" (the direct measurements, the "leaves"): do we want an operation to define that the simulation would draw from a distribution, so, do we want to implement a distribution as an operation ? I would say: Given that we want a distribution, as outcome, as part of a measurement, why can't we re-use this same distribution to "teach" the simulation engine on the "leaves"? If you would insist on an operation for that, ok, but then only when you allow for 0..* operations on a measure, otherwise we get that duplication-of-library-nightmare..   I agree that is should be possible to state a measurement with respect to a confidence interval. But that relationship is a relationship between measurements of different measures. Stating it as an attribute is a mistake. [hdm] We need to be practical. We need to care for maintainability and scalability. See above.    From: Henk de Man [mailto:hdman@cordys.com] Sent: Tuesday, July 17, 2012 8:01 AM To: Larry Hines Cc: Juergen Boldt; issues@omg.org; smm-rtf@omg.org; Henk de Man; Alain Picard; Pete Rivett; Arne Berre; fred.a.cummins Subject: Re: issues 17472 - 17475 -- SMM RTF issues  Larry,  See below. On Sat, Jul 14, 2012 at 8:40 PM, Larry Hines wrote: With respect to #17472: These attributes seem better fit for Measure, the method of measurement. A measurement is a single data point obtained by applying a measure. The measure may have precision, confidence bounds and a distribution. [hdm] ÂWith respect to #17472: It is not correct to state that a measurement is a single data point. See e.g.ÂHubbard, Douglas W., How to Measure Anything, Finding the Value of âIntangiblesâ in Business, John Wiley & Sons, New Jersey, 2010. His main point is, that, in business measurements, just point measures aren't good measures. Often the point measure is not known or is not 100 % trustable. It is very important to be able to state measurement (so the result of measurement) as something that resides in a confidence interval, according to a certain level of confidence. ÂA measurement can be a point value. But it should also be possible to state a measurement as e.g. a confidence interval, with a confidence value. So-far about confidence interval and confidence. Now about distribution (stochastic distribution): A Monte Carlo measurement (i.e. outcome of Multi Carlo simulation or experiment) is typically expressed as a set of values that can be best described by a distribution curve. Stochastic enabling of measurement is important. But the distribution itself can not very well be defined as part of the measure, because the measure is re-usable in different contexts, and the distribution will often be specific to that context.   With respect to #17473: Each of these types seem to be categories of measures. Perhaps SMM should include some built-in measure categories such as Estimators, Simulators and Benchmarks. [hdm] ÂWith respect to #17473: It is not about the measure, but about the observation. It is about a specification the type of the observation. Note that the same measure can be used for for estimated measurement, as-is measurement, simulated measurement.  With respect to #17474: Measures can be equated. If there are multiple, equally good direct measures for a given characteristic then they are be directly associated by the measure equivalence relationship. [hdm]ÂWith respect to #17474.ÂThis is not about equivalence. ÂIt is like method overloading in Java ... When the same direct measure is applied in different contexts, slightly different or additional arguments (parameters) might be required in its operation. So, actually, more than one operation would be required. During discussions with Alain Picard this came out as the best solution.  With respect to #17475: Iâm good with this.   From: Juergen Boldt [mailto:juergen@omg.org] Sent: Friday, July 13, 2012 10:59 AM To: issues@omg.org; smm-rtf@omg.org Subject: issues 17472 - 17475 -- SMM RTF issues  This is issue # 17472ÂÂÂFrom: Henk de Man From: Larry Hines To: Henk de Man CC: Juergen Boldt , "issues@omg.org" , "smm-rtf@omg.org" , Alain Picard , Pete Rivett , Arne Berre , "fred.a.cummins" Subject: RE: issues 17472 -- SMM RTF issues Thread-Topic: issues 17472 -- SMM RTF issues Thread-Index: AQHNZCP79dR81qHE+0yXusLumPpeXJcvfDUA///A8SCAAdUvgIABXpvQ Date: Fri, 20 Jul 2012 19:41:25 +0000 Accept-Language: en-US X-MS-Has-Attach: X-MS-TNEF-Correlator: x-originating-ip: [10.24.11.82] X-OriginalArrivalTime: 20 Jul 2012 19:41:26.0662 (UTC) FILETIME=[A6C17A60:01CD66AF] This is a very good idea. Do we know what the distribution type is? Normal for example. Or maybe Poisson? Normal just requires the mean and standard deviation. Poisson requires the mean (often called lambda). Other distributions? There are many. But just those two may be enough for most circumstances in VDML. From: Henk de Man [mailto:hdman@cordys.com] Sent: Thursday, July 19, 2012 11:30 AM To: Larry Hines Cc: Juergen Boldt; issues@omg.org; smm-rtf@omg.org; Alain Picard; Pete Rivett; Arne Berre; fred.a.cummins; Henk de Man Subject: Re: issues 17472 -- SMM RTF issues Larry, See attachment for detailed discussion on your topic below. I think I came into your direction, but this would require a couple of things to be clarified, see attachment. Can you react to it and see what you think ? Regards, Henk de Man On Wed, Jul 18, 2012 at 9:30 PM, Larry Hines wrote: Henk, A measure is an evaluation process that assigns a comparable numeric or symbolic value to an entity in order to characterize a selected quantity or trait of the entity. The evaluation process may be a mathematically formula applied to given values, the application of a tool to an entity or the fuzzy logic of a business analyst. It can be any process which evaluates an entity to produce a comparable value. Such evaluations are a large part of what business analyst are paid for. In fact, if the business analyst process is not defined (at least by a name) as a measure in SMM then his or her measurements cannot be represented in SMM. Width is inherently part of a line. It.s value is obtained by measuring the line. Precision and accuracy are likewise obtained by evaluation processes. Precision: Precision is the degree to which the correctness of a measurement is expressed. Often the precision of a measure is identical to its unit of measure. Examining the ruler that I keep in my desk I see its precision is one tenth of an inch on one side and a millimeter of the other. One tenth of an inch is a measurement inherent to measuring with that side of my ruler. Since precision is not always identical unit of measure, adding precision as an optional attribute of dimensional measure. Accuracy: Based up my experience I would evaluate the accuracy of Google Maps as very good. .Very good. is better than .good. but worse than .perfect.. If one wants a numeric value for the accuracy of Google Maps, we could select a number of places for which we have known longitudes and latitudes and we test Google Maps. The average of the absolute divergence from the known location would be one measure of Google Map.s accuracy. A range over a linear space is given by two values. One can compare the minimums, maximums, midpoints and lengths of ranges. But the ranges themselves are not directly comparable. That is, given ranges R1 and R2, R1 < R2 is not determined. Consequently, ranges are not measurements. One can, however, find a lower bound and a upper bound. That.s two measurements. The two measurements have the same measurand and are under the same observation. Yes, there is some inconvenience here. But what you want to represent can be represented correctly in SMM. With respect to scalability, if all one wants is the top level 100 measurements without any of the myriad of measurement aggregated to determine them, then that is all that SMM requires. The lower level measurements are not required. One can cut it off at any level. If one wants the top 100 distributions then one only needs the 100 sets of measurements that imply those 100 distributions, not the million measurements that went into their calculation. My basic point is that we should not lose the definitions of measure and measurement. If we lose those definitions then we have nothing. I.m tempted to propose an interval measure but that.s a can of worms to be thought out carefully. From: Henk de Man [mailto:hdman@cordys.com] Sent: Wednesday, July 18, 2012 11:17 AM To: Larry Hines Cc: Juergen Boldt; issues@omg.org; smm-rtf@omg.org; Alain Picard; Pete Rivett; Arne Berre; fred.a.cummins; Henk de Man Subject: Re: issues 17472 -- SMM RTF issues Larry, See below. On Tue, Jul 17, 2012 at 3:56 PM, Larry Hines wrote: Where does the value of .certain level of confidence. come from? [hdm] It is as good as the business analyst can make it. Often he/she doesn't have a point value, but just knows that e.g. the value is between "this" and "that" with x % probability ("confidence"). In other words: there's a degree of uncertainty left, and the only thing the measure does is narrowing down the uncertainty to the bounds of the confidence interval. But note again that these bounds aren't coming by measure (library), but are outcome of the mind of the business analyst on the spot. See the many examples in Hubbard, Douglas W., How to Measure Anything, Finding the Value of .Intangibles. in Business, John Wiley & Sons, New Jersey, 2010. It comes from a measure of confidence which is a distinct measure from the original measure. [hdm] No, no. See above. Precision of a measurement is not given by the measure.. It is inherently or intrinsically part of the measurement itself. This is important. The certain level of confidence is a measurement which is distinct from the original measurement. [hdm] No, because in many of these situations there is no "original measurement". There are just these bounds. And these form the measurement. Don't make each bound a separate measurement, because then we get burdened with so many objects in the model. And note that for all of these things the measure is just the measure: its name, its characteristic that is measures, its unit, etc. Care for maintainability and purity and re-usability of the library. They are indeed related measurements and it would be nice if SMM or another meta-model captured that relationship. We inserted the demonstration of trending data in SMM to answer a similar issue. Often the trend of the measurements is more important that the absolute value of the individual measurements. But that doesn.t change what a measurement is, namely a single data point. [hdm] The proposal is to add e.g. confidence interval bounds and probability to the point value, to have the measurement being expressed by a couple of things, rather than just that point value. And note again: often a point value cannot be given, but confidence interval can be. See again Hubbard, who is very major on this. If I were to say that it.s 100 degrees F give or take 2 degrees F, I would be asserting two measurements, 100 degrees F and 2 degrees F, and I would be asserting a relationship between the two measurements. The same applies to a distribution of measurements. [hdm] Note that what we primarily want to achieve with the distribution is: an expression of outcome of measurement. For instance, in a MonteCarlo simulation, there are 10.000 (e.g.) measurements per each "leave" measure (the direct ones). All of these are aggregated up (so to say), taking into account the so many binary, rescale and collective effects of the other measures, and then "at the top", all these 10.000 measurements come out (per each "top" measure). Consider a VDML model with, say 100 measured characterstics. A simulation engine would generate (in its run-time database or memory) 100 * 10.000 = 1.000.000 measurements.. What we want to import back in the VDML model (the model of the business system or business system scenario), is not 1.000.000 measurements. Of course not. We want an observation for the entire simulation, containing then just 100 measurements, whereby each measurement has a distribution (based on curve fitting, for instance), that describes the outcome. That's also what Monte Carlo does: rendering distributions of simulated values of such characteristics. So, the distributions, as part of measurement, is not meant as direct measure operation, but primarily as measurement outcome. Describing a distribution of values, rather than just one point value. And what you say, storing the all the so-many from-simulation-memory values as separate measurement objects, would not be feasible, as it would load a VDML model with 1.000.000 objects... Now, given that we need a refinement of a measurement with distribution, we can further argue about what this means at the "bottom" (the direct measurements, the "leaves"): do we want an operation to define that the simulation would draw from a distribution, so, do we want to implement a distribution as an operation ? I would say: Given that we want a distribution, as outcome, as part of a measurement, why can't we re-use this same distribution to "teach" the simulation engine on the "leaves"? If you would insist on an operation for that, ok, but then only when you allow for 0..* operations on a measure, otherwise we get that duplication-of-library-nightmare.. I agree that is should be possible to state a measurement with respect to a confidence interval. But that relationship is a relationship between measurements of different measures. Stating it as an attribute is a mistake. [hdm] We need to be practical. We need to care for maintainability and scalability. See above. From: Henk de Man [mailto:hdman@cordys.com] Sent: Tuesday, July 17, 2012 8:01 AM To: Larry Hines Cc: Juergen Boldt; issues@omg.org; smm-rtf@omg.org; Henk de Man; Alain Picard; Pete Rivett; Arne Berre; fred.a.cummins Subject: Re: issues 17472 - 17475 -- SMM RTF issues Larry, See below. On Sat, Jul 14, 2012 at 8:40 PM, Larry Hines wrote: With respect to #17472: These attributes seem better fit for Measure, the method of measurement. A measurement is a single data point obtained by applying a measure. The measure may have precision, confidence bounds and a distribution. [hdm] With respect to #17472: It is not correct to state that a measurement is a single data point. See e.g. Hubbard, Douglas W., How to Measure Anything, Finding the Value of .Intangibles. in Business, John Wiley & Sons, New Jersey, 2010. His main point is, that, in business measurements, just point measures aren't good measures. Often the point measure is not known or is not 100 % trustable. It is very important to be able to state measurement (so the result of measurement) as something that resides in a confidence interval, according to a certain level of confidence. A measurement can be a point value. But it should also be possible to state a measurement as e.g. a confidence interval, with a confidence value. So-far about confidence interval and confidence. Now about distribution (stochastic distribution): A Monte Carlo measurement (i.e. outcome of Multi Carlo simulation or experiment) is typically expressed as a set of values that can be best described by a distribution curve. Stochastic enabling of measurement is important. But the distribution itself can not very well be defined as part of the measure, because the measure is re-usable in different contexts, and the distribution will often be specific to that context. With respect to #17473: Each of these types seem to be categories of measures. Perhaps SMM should include some built-in measure categories such as Estimators, Simulators and Benchmarks. [hdm] With respect to #17473: It is not about the measure, but about the observation. It is about a specification the type of the observation. Note that the same measure can be used for for estimated measurement, as-is measurement, simulated measurement. With respect to #17474: Measures can be equated. If there are multiple, equally good direct measures for a given characteristic then they are be directly associated by the measure equivalence relationship. [hdm] With respect to #17474. This is not about equivalence. It is like method overloading in Java ... When the same direct measure is applied in different contexts, slightly different or additional arguments (parameters) might be required in its operation. So, actually, more than one operation would be required. During discussions with Alain Picard this came out as the best solution. With respect to #17475: I.m good with this. From: Juergen Boldt [mailto:juergen@omg.org] Sent: Friday, July 13, 2012 10:59 AM To: issues@omg.org; smm-rtf@omg.org Subject: issues 17472 - 17475 -- SMM RTF issues This is issue # 17472 From: Henk de Man This message has been scanned by MailController. -- Henk de Man Research Director hdman@cordys.com www.cordys.com T +31 (0)341 37 5541 . M +31 (0)6 51 43 09 45 CORDYS . Improvinng Business Operations This message has been scanned by MailController. -- Henk de Man Research Director hdman@cordys.com www.cordys.com T +31 (0)341 37 5541 . M +31 (0)6 51 43 09 45 CORDYS . Improving Business Operations This message has been scanned by MailController. X-Google-DKIM-Signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=google.com; s=20120113; h=mime-version:in-reply-to:references:date:message-id:subject:from:to :cc:content-type:x-gm-message-state; bh=DL1VGAymbwwgwlBIGZaHU3wRZBgirgjIfCF0oXDzj7o=; b=oROY75o4TmYiYb7RtOhtyay26vmmGew7Chvl+8K2ajKIwyXG9/c2YwIFiHTZReknRS 8QhJBNKgd+y8TLQX0yKvyy8bAQBJP7p6BAFl62MXAe38cRUuStUzmSGlU65Ew3g7Mt+h 9VhAVgW3lufINwi49pR4ZTccxfeR4cL1mPhWAPK1Oa7/hFxapQmhA0tMMntteTvPkhSH CCOvuTeXFj4biqV5HaBYDe2uvkHgw4CLhQuyiZJ2CcpZraRPFPTjRshsyw8RjBHOlaJ2 cAl84pYz2Pm5ehj/oQVGjk+4MqYbhoEEYWWA0HM8y2G7U6fu5PSC5f4KKBZwlDyGlt/A Otzg== Date: Sat, 21 Jul 2012 18:29:12 +0200 Subject: Re: issues 17472 -- SMM RTF issues From: Henk de Man To: Larry Hines Cc: Juergen Boldt , "issues@omg.org" , "smm-rtf@omg.org" , Alain Picard , Pete Rivett , Arne Berre , "fred.a.cummins" , Henk de Man X-Gm-Message-State: ALoCoQkKXeB6pg5cDo+UxY1aQcMHm0oVfCc9NZZQgb8omXiAIBbewWMvuKf4+ZjYn+/kWkNfox32 Larry, The fundamental question is: if a "point value measurement" is not the "full truth", and we want to express "precision" of that measurement also, e.g. by a set of related collective measurements that "summarizes" ("condensus") the "full sample" (the "10.000 measurements"..), which minimal set of such measurements will then best describe the precision, and will best enable reproducing the distribution shape or histogram ? Is this the set of standard deviation, skewness and curtosis (assuming that the point value itself represents the mean) ? Or is there a more extendable way ? But is there any good and well-known algorithm that can reproduce the distribution or histogram based on these four "moments" ? So, I agree that it would be useful to "summarize" the "big underlying set of measurements" by just a minimal set of additional collective measurements. But what about these questions? What do you think ? See some links, for convenience. http://en.wikipedia.org/wiki/Moment_(mathematics) and http://www.pertrac.com/resources/investment-statistics-guide/assessing-skewness-and-kurtosis-in-the-return-distribution/ âMomentsâ: mean, standard deviation, skewness, kurtosis == http://stat-design.blogspot.nl/2011/06/using-mean-standard-deviation-skewness.html âvisual curve fittingâ based on the four moments == http://www.mathworks.nl/help/toolbox/stats/br5k833-1.html Finding distributions based on Pearson or Johnson, based on 4 moments (see also http://www.mathworks.nl/help/toolbox/stats/pearsrnd.html and http://www.mathworks.nl/help/toolbox/stats/johnsrnd.html Â) == Maybe also: http://www.stata.com/statalist/archive/2006-11/msg00075.html == http://stackoverflow.com/questions/4807398/how-to-generate-distributions-given-mean-sd-skew-and-kurtosis-in-r ((Note again that I thought to store the distribution signature itself, to avoid any difficulty on reproducing from such moments..)) Looking forward to your thoughts. Regards, Henk de Man On Fri, Jul 20, 2012 at 9:41 PM, Larry Hines wrote: This is a very good idea. Do we know what the distribution type is? Normal for example. Or maybe Poisson?  Normal just requires the mean and standard deviation. Poisson requires the mean (often called lambda).  Other distributions? There are many. But just those two may be enough for most circumstances in VDML.  From: Henk de Man [mailto:hdman@cordys.com] Sent: Thursday, July 19, 2012 11:30 AM To: Larry Hines Cc: Juergen Boldt; issues@omg.org; smm-rtf@omg.org; Alain Picard; Pete Rivett; Arne Berre; fred.a.cummins; Henk de Man Subject: Re: issues 17472 -- SMM RTF issues  Larry,  See attachment for detailed discussion on your topic below. I think I came into your direction, but this would require a couple of things to be clarified, see attachment. Can you react to it and see what you think ?  Regards,  Henk de Man  On Wed, Jul 18, 2012 at 9:30 PM, Larry Hines wrote: Henk,  A measure is an evaluation process that assigns a comparable numeric or symbolic value to an entity in order to characterize a selected quantity or trait of the entity.  The evaluation process may be a mathematically formula applied to given values, the application of a tool to an entity or the fuzzy logic of a business analyst. It can be any process which evaluates an entity to produce a comparable value. Such evaluations are a large part of what business analyst are paid for. In fact, if the business analyst process is not defined (at least by a name) as a measure in SMM then his or her measurements cannot be represented in SMM.  Width is inherently part of a line. Itâs value is obtained by measuring the line. Precision and accuracy are likewise obtained by evaluation processes.  Precision: Precision is the degree to which the correctness of a measurement is expressed. Often the precision of a measure is identical to its unit of measure. Examining the ruler that I keep in my desk I see its precision is one tenth of an inch on one side and a millimeter of the other. One tenth of an inch is a measurement inherent to measuring with that side of my ruler. Since precision is not always identical unit of measure, adding precision as an optional attribute of dimensional measure.  Accuracy: Based up my experience I would evaluate the accuracy of Google Maps as very good. âVery goodâ is better than âgoodâ but worse than âperfectâ.  If one wants a numeric value for the accuracy of Google Maps, we could select a number of places for which we have known longitudes and latitudes and we test Google Maps. The average of the absolute divergence from the known location would be one measure of Google Mapâs accuracy.  A range over a linear space is given by two values. One can compare the minimums, maximums, midpoints and lengths of ranges. But the ranges themselves are not directly comparable. That is, given ranges R1 and R2, R1 < R2 is not determined. Consequently, ranges are not measurements.  One can, however, find a lower bound and a upper bound. Thatâs two measurements. The two measurements have the same measurand and are under the same observation. Yes, there is some inconvenience here. But what you want to represent can be represented correctly in SMM. With respect to scalability, if all one wants is the top level 100 measurements without any of the myriad of measurement aggregated to determine them, then that is all that SMM requires. ÂÂÂThe lower level measurements are not required. One can cut it off at any level.  If one wants the top 100 distributions then one only needs the 100 sets of measurements that imply those 100 distributions, not the million measurements that went into their calculation.  My basic point is that we should not lose the definitions of measure and measurement. If we lose those definitions then we have nothing.  Iâm tempted to propose an interval measure but thatâs a can of worms to be thought out carefully.  From: Henk de Man [mailto:hdman@cordys.com] Sent: Wednesday, July 18, 2012 11:17 AM To: Larry Hines Cc: Juergen Boldt; issues@omg.org; smm-rtf@omg.org; Alain Picard; Pete Rivett; Arne Berre; fred.a.cummins; Henk de Man Subject: Re: issues 17472 -- SMM RTF issues  Larry,  See below. On Tue, Jul 17, 2012 at 3:56 PM, Larry Hines wrote: Where does the value of âcertain level of confidenceâ come from? [hdm] It is as good as the business analyst can make it. Often he/she doesn't have a point value, but just knows that e.g. the value is between "this" and "that" with x % probability ("confidence"). In other words: there's a degree of uncertainty left, and the only thing the measure does is narrowing down the uncertainty to the bounds of the confidence interval. But note again that these bounds aren't coming by measure (library), but are outcome of the mind of the business analyst on the spot. See the many examples inÂHubbard, Douglas W.,ÂHow to Measure Anything, Finding the Value of âIntangiblesâ in Business, John Wiley & Sons, New Jersey, 2010.  It comes from a measure of confidence which is a distinct measure from the original measure.  [hdm] No, no. See above. ÂPrecision of a measurement is not given by the measure.. It is inherently or intrinsically part of the measurement itself. This is important. The certain level of confidence is a measurement which is distinct from the original measurement. [hdm] No, because in many of these situations there is no "original measurement". There are just these bounds. And these form the measurement. Don't make each bound a separate measurement, because then we get burdened with so Âmany objects in the model. And note that for all of these things the measure is just the measure: its name, its characteristic that is measures, its unit, etc. Care for maintainability and purity and re-usability of the library.   They are indeed related measurements and it would be nice if SMM or another meta-model captured that relationship.  We inserted the demonstration of trending data in SMM to answer a similar issue. Often the trend of the measurements is more important that the absolute value of the individual measurements. But that doesnât change what a measurement is, namely a single data point. [hdm] The proposal is to add e.g. confidence interval bounds and probability to the point value, to have the measurement being expressed by a couple of things, rather than just that point value. And note again: often a point value cannot be given, but confidence interval can be. See again Hubbard, who is very major on this.   If I were to say that itâs 100 degrees F give or take 2 degrees F, I would be asserting two measurements, 100 degrees F and 2 degrees F, and I would be asserting a relationship between the two measurements. The same applies to a distribution of measurements. [hdm] Note that what we primarily want to achieve with the distribution is: an expression of outcome of measurement. For instance, in a MonteCarlo simulation, there are 10.000 (e.g.) measurements per each "leave" measure (the direct ones). All of these are aggregated up (so to say), taking into account the so many binary, rescale and collective effects of the other measures, and then "at the top", all these 10.000 measurements come out (per each "top" measure). Consider a VDML model with, say 100 measured characterstics. A simulation engine would generate (in its run-time database or memory) 100 * 10.000 = 1.000.000 measurements.. What we want to import back in the VDML model (the model of the business system or business system scenario), is not 1.000.000 measurements. Of course not. We want an observation for the entire simulation, containing then just 100 measurements, whereby each measurement has a distribution (based on curve fitting, for instance), that describes the outcome. That's also what Monte Carlo does: rendering distributions of simulated values of such characteristics. So, the distributions, as part of measurement, is not meant as direct measure operation, but primarily as measurement outcome. Describing a distribution of values, rather than just one point value. And what you say, storing the all the so-many from-simulation-memory values as separate measurement objects, would not be feasible, as it would load a VDML model with 1.000.000 objects...  Now, given that we need a refinement of a measurement with distribution, we can further argue about what this means at the "bottom" (the direct measurements, the "leaves"): do we want an operation to define that the simulation would draw from a distribution, so, do we want to implement a distribution as an operation ? I would say: Given that we want a distribution, as outcome, as part of a measurement, why can't we re-use this same distribution to "teach" the simulation engine on the "leaves"? If you would insist on an operation for that, ok, but then only when you allow for 0..* operations on a measure, otherwise we get that duplication-of-library-nightmare..   I agree that is should be possible to state a measurement with respect to a confidence interval. But that relationship is a relationship between measurements of different measures. Stating it as an attribute is a mistake. [hdm] We need to be practical. We need to care for maintainability and scalability. See above.    From: Henk de Man [mailto:hdman@cordys.com] Sent: Tuesday, July 17, 2012 8:01 AM To: Larry Hines Cc: Juergen Boldt; issues@omg.org; smm-rtf@omg.org; Henk de Man; Alain Picard; Pete Rivett; Arne Berre; fred.a.cummins Subject: Re: issues 17472 - 17475 -- SMM RTF issues  Larry,  See below. On Sat, Jul 14, 2012 at 8:40 PM, Larry Hines wrote: With respect to #17472: These attributes seem better fit for Measure, the method of measurement. A measurement is a single data point obtained by applying a measure. The measure may have precision, confidence bounds and a distribution. [hdm] ÂWith respect to #17472: It is not correct to state that a measurement is a single data point. See e.g.ÂHubbard, Douglas W., How to Measure Anything, Finding the Value of âIntangiblesâ in Business, John Wiley & Sons, New Jersey, 2010. His main point is, that, in business measurements, just point measures aren't good measures. Often the point measure is not known or is not 100 % trustable. It is very important to be able to state measurement (so the result of measurement) as something that resides in a confidence interval, according to a certain level of confidence. ÂA measurement can be a point value. But it should also be possible to state a measurement as e.g. a confidence interval, with a confidence value. So-far about confidence interval and confidence. Now about distribution (stochastic distribution): A Monte Carlo measurement (i.e. outcome of Multi Carlo simulation or experiment) is typically expressed as a set of values that can be best described by a distribution curve. Stochastic enabling of measurement is important. But the distribution itself can not very well be defined as part of the measure, because the measure is re-usable in different contexts, and the distribution will often be specific to that context.   With respect to #17473: Each of these types seem to be categories of measures. Perhaps SMM should include some built-in measure categories such as Estimators, Simulators and Benchmarks. [hdm] ÂWith respect to #17473: It is not about the measure, but about the observation. It is about a specification the type of the observation. Note that the same measure can be used for for estimated measurement, as-is measurement, simulated measurement.  With respect to #17474: Measures can be equated. If there are multiple, equally good direct measures for a given characteristic then they are be directly associated by the measure equivalence relationship. [hdm]ÂWith respect to #17474.ÂThis is not about equivalence. ÂIt is like method overloading in Java ... When the same direct measure is applied in different contexts, slightly different or additional arguments (parameters) might be required in its operation. So, actually, more than one operation would be required. During discussions with Alain Picard this came out as the best solution.  With respect to #17475: Iâm good with this.   From: Juergen Boldt [mailto:juergen@omg.org] Sent: Friday, July 13, 2012 10:59 AM To: issues@omg.org; smm-rtf@omg.org Subject: issues 17472 - 17475 -- SMM RTF issues  This is issue # 17472ÂÂÂFrom: Henk de Man