I've got my fingers crossed that at least one of the designs soon to be released has a geometry that will make me happy.
I would be surprised if Goddard's knives were unsuitable. If you regrind those then you are likely to regrind everything. I think an ideal geometry would be similar to Krein's regrinds of the Spyderco's that have been propogating as of late.
I don't want any lies completely ignored.
Was not referring to you with that statement.
I think (but don't know because you haven't explained it clearly) that one of your concerns about the data is that there is too much uncertainty in the model parameters. If so (and I recognize I'm taking a stretch here), how does that square with the idea that the model "fits their data perfectly"?
The reason it doesn't fit is because of calculations which as noted introduce another nonlinearity and amplify the noise. It fits the actual raw data fine as noted, this would then obviously prove that the physical behavior is modeled. You can take any data set and by repeated calculations make it unable to be modeled by anything becuase you can make it undetermined as the noise will exceed the signal by sufficient amounts. This is a trivial pure math proof, as it is just an infinite limit problem. If you don't see this then I can generate a simple data set for some simple physical law and then show how it can be made undefined. One trivial way is to just look at differences because very quickly the signal to noise will approach zero. This of course does not mean the data does not fit the model as again you have to look at the raw data, most times, in some cases you can linearize it, or put it through some kind of filter/window to remove background noise, but such things are highly suspect to debate. In general it is best to work with the raw data and add any extra functions to model whatever you are trying to remove directly. This also allows you to determine any correlation functions associated with the unwanted behavior.
Could you please provide the nonlinear math problem to be solved using monte carlo simulation?
I have discussed this in detail in other posts where I show how you can calculate the spread of the cut ratio using a monte carlo simulation on the raw data. This in fact ignores all models, it doesn't even use one and uses the raw data directly to calculate the cut ratio. To be clear, there is no model involved at all. The performance advantage comes from the unaltered data. I have done that in the past on many occasions.
You can do the exact same thing for the shaving sharpness comparison. If you are actually serious about wanting the calculations performed for the REST data then I'll do them shortly. I have to do a series of ANOVA calculations on a large scale sharpness project first though for someone else who is working on a huge edge retention project and wants proof one way or another of some conclusions.
I can write some code for you to actually do this yourself if you tell me which system you are using and which compilers you have. I actually do it in free software, out of amusement I wrote the montecarlo part in AWK last year, but sensibly it would be in matlab or similar. There are free versions of this you can use which would be much more robust and flexible. If you tell me what kind of interface/options you would want it would be easy to generate the code.
Would you please list the assumptions necessary that are not true in this case?
if you want to compare averages then your data would correctly have to be modeled by F(X)=c because an average is just a least squares fit of that model. If this isn't the model which describes your data then you can't use the parameters it predicts obviously.
But it looks to me like at the angles we tested, FFD2 outperformed s90V very strongly.
I am not sure you understood what I said, I was noting where the difference in performance would be greater for your process and thus show the largest difference in your favor. If you wanted to be complete you could show the worst comparison in your favor which would be high angles slicing with S90V with coarse edges. The user could then infer in general to expect a difference inbetween those two. In fact if you wanted you could do the results over a range of angles, cut types and grits and send me the data and I would model all of that and allow you (or the user) to calculate what advantage he would see given his choices in the above. I have been meaning to do that personally for some time but generating the data does take awhile and I find it rather boring considering the volume of material that the steels I tend to favor now require to blunt, M2 at 66 HRC for example.
I'm not sure which example you refer to here. Could you give me a specific citation?
Buck's Ionfusion blades are a known commercial example. There is also some data going to be released soon of another blade material which actually ignores CATRA testing as noted, the blade does not actually blunt at all on a full dat on the machine. Before anyone gets excited the blade material is useless for humans, it just exploits a huge systematic bias in such testing as did the Ionfusion blades.
I'm not clear what you consider the required methods to produce unbiased results, so please tell me.
The user and the test organizer can not know which blades have undergone which process. This is the way I conduct the test group that I organize for example. I have also done this myself and only after I have published the work to the maker will I know which blades have which steel/heat treatment. I also do other things to remove bias when I do know the blades details such as have an unknown amount of work done by another and only after i have competed my work is this revealed. This prevents me from introducing any bias due to personal preference because I have no ability to know when the performance should degrade obviously. There are other things you can do such as use uncalibrated stock which is only calibrated after the cutting, etc. . I do that as well.
-Cliff
